model rankings

Best AI Models

The definitive, opinionated ranking of the best AI models in 2026 — by consensus across benchmarks, reviews and real-world testing. Updated monthly.

Updated October 2026

Quick answer: Claude Opus 5.5 is the best AI model available: it has the highest score Artificial Analysis has measured on its Intelligence Index (58), at $4/$20 per million tokens. Claude Sonnet 5.5 is second (56) at half the token price, and Meta’s Muse Wonder Full matches that score but is sold only inside Muse Code, with no API. GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon follow on 53, and GPT-6.1 Sol is the best value at $0.72 per completed task. The best open-weight model is DeepSeek V4.1-Flash, with GLM-5.3 and Kimi K3 the highest-scoring open models. The full ranking, the evidence behind each placement and our picks by job are below.

This is an opinionated ranking. It is not a copy of any single leaderboard. We rank by consensus — across benchmarks (vendor and standardized), independent reviews, public arenas and our own real-world testing — and we cite the evidence behind every placement. Where a figure isn’t verifiable, we say so rather than invent one.

Why this ranking moves every month

AI rankings are not stable. Three forces keep the order in motion.

New models ship constantly. The major labs release new flagships, point releases and price cuts every few weeks, and open-weight models follow close behind. A ranking that is three months old is already out of date.

The benchmarks themselves rotate. Benchmarks saturate, leak into training data and are replaced by harder ones, and independent indices re-base their scales when that happens. The number everyone quoted last quarter is often not the one that matters this quarter.

The same model scores differently depending on who runs it. A vendor’s own harness usually reports a higher score than a standardised leaderboard does for the same model on the same benchmark. Both numbers are real; the difference is the harness, not the model. We treat vendor numbers as a ceiling and standardised leaderboards as a floor, and say which is which.

So this page is re-scored every month, and whenever a major launch or price change demands it. It is a judgement backed by the evidence below, not a copy of any single leaderboard.

Capability per dollar

One picture before the table. Every generally-available model we score, plotted by The AI Rankings score — our coding-weighted composite, the same judgement behind the ranking below — against blended API price. The black line is the value frontier: at each price, the most capable model money can buy, so anything inside the line is beaten on both price and capability. Note the distinction: the value frontier is about cost-efficiency, not capability class — a small model like Claude Haiku 5.5 sits on it simply because nothing cheaper beats it, while the board’s Frontier tier (the capability class) is a different thing, shown in the table below and in the chart data. Re-scored monthly alongside the ranking, and whenever a major launch or price move demands it; every movement lands in the changelog.

The ranking: top 58 AI models

Ranked by overall capability and consensus standing, weighted toward repository-scale coding and agentic work (where the most comparable cross-model data exists), then reasoning, knowledge work and independent sentiment. Tiers, not just rank, are the point: a model’s tier tells you more than its exact position.

Reading the SWE-bench column: ᵖᵛ = SWE-bench Pro, vendor harness (a ceiling); ᵖˢ = SWE-bench Pro, Scale standardized harness (a floor); ᵛ = SWE-bench Verified (older, Python-only, quoted at launch). Where two figures appear, the first is the harness the provider reports. n/a means we don’t yet have a verified figure for that cell — usually because the model’s own page is still in build — not that the capability doesn’t exist; see the linked model page for detail as it lands.

Two structural facts stand out. Anthropic holds five of the top ten — Opus 5.5 and Sonnet 5.5, then Mythos 5.1 and Fable 5.1 at #4 and #5 and Mythos 5 at #10 — Meta has two, Muse Wonder Full at #3 and Muse Wonder at #9, OpenAI has two, GPT-6 Astra at #6 and GPT-6.1 Sol at #8, and Google has one: Gemini 4 Argon at #7. GPT-5.6 Sol, Grok 4.7, Grok 4.6, GLM-5.3 and Kimi K3 packed tightly at #14–18, and the open-weight tier has closed the gap: DeepSeek V4 (80.6%) and MiniMax M3 (80.5%) match Gemini 3.1 Pro on SWE-bench Verified, with Kimi K2.6 (80.2%) just behind — at a fraction of the price. (Alibaba’s Qwen3.8-Max now sits at #19 on an independent score of 45; see the note below.) The frontier is usable again — Fable 5 came back to general availability on 1 July after its 18-day export-control suspension — while the floor rises fast; only the two Mythos models stay out of general reach, restricted to vetted organisations. At #12: Claude Opus 5 (24 July 2026) led the board’s headline independent metric until Fable 5.1 landed — 61 on the Artificial Analysis Intelligence Index at max effort, the highest figure it had ever recorded at the 24 July launch, one point above Fable 5 (then 60) and two above GPT-5.6 Sol (then 59) — and does it at half Fable 5’s price ($5/$25). (AA’s August re-run rescored the cluster upward — Opus 5 max 63, Fable 5 max 62, Grok 4.6 and Sol max 61 — with the order unchanged; Fable 5.1 then cleared all of them at 66 on 1 September. On the re-based v4.3 index of 7 September, which is harder across the board, Opus 5 reads 51 — third, behind Fable 5.1 and GPT-6 Astra on 53 — at $5.86 per completed task.) It also tops SWE-bench Verified (96.0% on Anthropic’s harness, 97.0% on vals.ai’s) and posts the largest ARC-AGI-3 result yet (30.16%, roughly 4x GPT-5.6 Sol). We hold it at #12 rather than higher for three honest reasons: GPT-6 Astra and GPT-6.1 Sol now out-score it on the re-based index (53 and 52 to 51) at far lower costs per task ($3.26 and $0.72 to $5.86); Anthropic’s own System Card states it is “not more capable overall” than Fable 5, which keeps a narrow SWE-bench Pro lead (80.3% vs 79.2%) under our coding-weighted methodology; and it hallucinates slightly more than Opus 4.8 (AA-Omniscience rate ~50%, up ~14 points). But it clearly clears GPT-5.6 Sol and Kimi K3 beneath it.

New at #1: Claude Opus 5.5 (22 September 2026) takes the top of the board on the strongest independent evidence available. Artificial Analysis scored it 58 on its Intelligence Index v4.3 at max effort — the highest score it has recorded, five points clear of GPT-6 Astra and Fable 5.1 (53) — leading six of the index’s ten evaluations and trailing on three (CritPt, AA-LCR, GDP.pdf). On Anthropic’s own table it beats Fable 5.1 on Terminal-Bench 4.0 (66.4% to 55.8%) and FrontierCode (54.4% to 50.3%), which is what moves it above the restricted Mythos 5.1 (60.9% on the same vendor harness): it is both higher-scoring and generally available. It costs $4/$20, a fifth below Opus 5, with cache reads down 60 per cent to $0.20. Three caveats keep the entry honest. It publishes no SWE-bench figure of any kind, where Opus 5 reported 79.2% Pro and 96.0% Verified. It spends about 119,000 output tokens per index task against Opus 5’s 73,000, so the price cut leaves cost per completed task essentially unchanged at $5.98. And Artificial Analysis’s own Terminal-Bench 4.0 run reads 59.6%, level with Astra and seven points under Anthropic’s figure. Biology and cybersecurity requests that trip Anthropic’s safeguards are answered by Opus 4.8 unless you are in one of its verification programmes.

New at #2: Claude Sonnet 5.5 (28 September 2026) enters on the same kind of evidence that put Opus 5.5 at #1. Artificial Analysis scored it 56 on its Intelligence Index at max effort — 18 points above Sonnet 5, two under Opus 5.5 (58) and three above Fable 5.1 and GPT-6 Astra (53) — with Terminal-Bench 4.0 at 64% on its own run, above Opus 5.5 and Astra (about 60%), and parity with Opus 5.5 on AA-Briefcase, GDPval-AA and AutomationBench-AA. On Anthropic’s table it posts 70.6% on Terminal-Bench 4.0 against the restricted Mythos 5.1’s 60.9% on the same vendor harness, which is what moves it above Mythos 5.1: higher-scoring and generally available, at $2/$10, unchanged from Sonnet 5. Four things keep it below Opus 5.5. It is two points lower on the index, and less accurate on factual questions (54% to 66% on AA-Omniscience), with a gap of about six points on Humanity’s Last Exam and SciCode. At max effort it uses about 193,000 output tokens per index task, the most Artificial Analysis has measured, so a completed task costs $7.60 — more than Opus 5.5’s $5.98 at twice the token price, and about 50% more than Sonnet 5, against Anthropic’s own claim of up to 30% less per task. Its API default is high effort, where the index reads 47, not 56. And it publishes no SWE-bench figure of any kind, where Sonnet 5 reported 63.2% Pro. High-risk cybersecurity tasks fall back to Sonnet 5 unless you are in Anthropic’s Cyber Verification Program.

New at #3: Muse Wonder Full (12 October 2026) is the stronger of the two models Meta released together. Artificial Analysis scored it 56 on the Intelligence Index v4.3.2 at max effort, level with Claude Sonnet 5.5 and two under Opus 5.5 (58), and three above Fable 5.1 and GPT-6 Astra (53). It sits below Sonnet 5.5 because a tie on the index does not displace the incumbent, and Sonnet 5.5 is on a public API at $2/$10 where Muse Wonder Full is available only to Muse Code subscribers, with no API and no price per token. Without a price it gets no point on the value chart and no pick. Meta published no SWE-bench figure, no parameter count and no account of how it differs from Muse Wonder, and on Meta’s own table it trails Opus 5.5 and Astra on Terminal-Bench 4.0 (57.1 against 59.6 and 59.1) while leading both on DeepSWE 1.1 (79.0 against 74.2 and 74.1).

New at #9: Muse Wonder (12 October 2026) is the Wonder model developers can buy: on the Meta Model API at $4/$20 per million tokens, the same token price as Opus 5.5. Artificial Analysis scored it 52 on the Intelligence Index v4.3.2 at max effort, level with GPT-6.1 Sol and above Opus 5 (51) and Fable 5 (50). It sits directly below GPT-6.1 Sol because a completed index task costs $4.52 against Sol’s $0.72, at twice the token price, and above Mythos 5, Fable 5 and Opus 5 on the index lead. Meta published no SWE-bench figure, no knowledge cutoff and no maximum output length, and its full safety report was not available at launch.

New at #7: Gemini 4 Argon (30 September 2026) is Google’s first new frontier model since Gemini 3.1 Pro in February, announced without Gemini 3.5 Pro having shipped. Artificial Analysis scored it 53 on the Intelligence Index v4.3.2 at high reasoning, level with GPT-6 Astra and Fable 5.1 and one point above GPT-6.1 Sol (52), at $1.99 per completed index task on an introductory $2/$10, with a 15% hallucination rate on AA-Omniscience, the lowest it has measured for a model scoring above 45, and first place on AutomationBench-AA (77.5%). Vals puts it first on the Vals Index at 68.9%, and on Google’s own table it leads DeepSWE v1.1 at 77.9% against Opus 5.5’s 74.2% and Astra’s 74.1%. It places below Astra because a tie on the index does not displace the incumbent and Astra leads it on Artificial Analysis’s Terminal-Bench 4.0 run (59.1% to 57.1%), and above GPT-6.1 Sol on the one-point index lead. Google first rolled it out to cyber defenders in its Fairwind Program and made it generally available on 13 October, in the Gemini app for Google AI Pro and Ultra subscribers and on the Gemini API for paid customers, so it now has a point on the value chart. It publishes no SWE-bench figure of any kind, the $2/$10 price doubles to $4/$20 after 30 November, and it used about 62,000 output tokens per index task against Astra’s 27,000.

New at #20: Claude Haiku 5.5 (7 October 2026) is Anthropic’s small model, generally available on the Claude API at $0.10/$0.50 per million tokens for prompts up to 100,000 tokens and $0.50/$2.50 above (Anthropic). Artificial Analysis scored it 43 on the Intelligence Index at max effort, at $0.21 per completed task: one and two points under Kimi K3 (44) and Qwen3.8-Max (45) above it, and ahead of Claude Opus 4.8 (42) and Gemini 3.8 Flash (41) below. Three things keep it from placing higher: Anthropic published no SWE-bench figure, its Terminal-Bench 4.0 result on Anthropic’s own table is 39.2% against Sonnet 5.5’s 70.6%, and the API’s default medium effort scores 34, nine points under the headline.

New at #30: Mistral Large 4 (6 October 2026) is Mistral’s 1-trillion-parameter model, in public preview on its API at $1.36/$4.18 per million tokens, with open weights promised by the end of October (Mistral). Artificial Analysis scored the preview 38 on the Intelligence Index v4.3.2, level with Claude Sonnet 5 and one point under Grok 4.5, GPT-5.5 and DeepSeek V4.1-Flash (39), at $1.13 per completed task. It sits directly below Sonnet 5 because Sonnet 5 is generally available and publishes SWE-bench figures, while Mistral published none. As a preview it gets no point on the value chart and no pick; the placement is reviewed when the weights and a licence are published.

At #8: GPT-6.1 Sol (29 September 2026) arrived a week after GPT-6 Sol at the same $2/$10, and this time the gain is capability. Artificial Analysis scored it 52 on the Intelligence Index v4.3 at max effort, four above GPT-6 Sol and one below GPT-6 Astra, with the largest gains on Terminal-Bench 4.0 (12 points) and AA-Omniscience accuracy (8), at $0.72 per completed index task against Astra’s $3.26. It places below Astra, which keeps a two-point lead on the Coding Agent Index at max effort, and Gemini 4 Argon (53), and above Mythos 5, Fable 5 and Opus 5 (51), which it out-scores on the index at a fraction of their cost per task. It sits no higher because it publishes no SWE-bench figure of any kind, its hallucination rate at max effort is still 54 per cent, and OpenAI’s own safety addendum reports more attempts to bypass security restrictions than Astra in some test scenarios.

At #13: GPT-6 Sol (22 September 2026) is a price move rather than a capability one, and it resets the value end of this board. Artificial Analysis scored it 48 on the Intelligence Index v4.3 — one point above GPT-5.6 Sol (47), two above Grok 4.7 — and 57 on its Coding Agent Index, up two. What changed is the bill: $2/$10 per million tokens against $4/$20, billed as a standard price rather than an introductory one, for $1.06 per completed index task against GPT-5.6 Sol’s $1.99 and Grok 4.6’s $1.86 — the lowest cost per completed task on this board outside the small tier until GPT-6.1 Sol followed it a week later at $0.72, which took the best-value slot from it. Its sibling GPT-6 Luna ($0.10/$0.50, index 37, $0.07 per task) is covered on the same page and is not separately ranked. It sits at #13, below its successor, because the capability gain is one index point, both models regress on GDPval-AA v2.1 knowledge work, there is no SWE-bench figure of any kind, and its hallucination rate, while down from 92 to 60 per cent at max effort, is still high in absolute terms.

At #4 and #5: Claude Mythos 5.1 and Claude Fable 5.1 (1 September 2026) are one model at two safeguard levels, and they sit side by side on different evidence. Fable 5.1, the public one, earns #5 on the strongest independent result on this board: Artificial Analysis scored it 66 on the Intelligence Index at launch, 1st of 192 models tracked and the first score above 65 the index had recorded, against Opus 5’s 63 and Fable 5’s 62 — and on the harder re-based v4.3 index of 7 September it reads 53, tied first with GPT-6 Astra, above Opus 5 (51) and Fable 5 (50), and still ahead of Astra on the knowledge-work (GDPval-AA v2 1764 to 1662), scientific-coding and factual-accuracy components. Anthropic’s own table backs it — Terminal-Bench-Science 0.1 at 52.6% against Fable 5’s 24.7% (more than double, and clear of Opus 5’s 29.0% and Sol’s 22.4%), AutomationBench 31.4% against 17.1%, OSWorld 2.0 77.9% against 72.9% — and token prices are unchanged at $10/$50 with cache reads cut 75% to $0.25, which Anthropic puts at roughly 25% off a typical workload and up to 45% off heavily agentic work. Three things keep it at #5 rather than #4 overall, and out of our best-overall pick entirely. It publishes no SWE-bench figure of any kind — neither Pro nor Verified — which costs it under our coding-weighted methodology and leaves the board’s highest SWE-bench Pro score in the hands of its own predecessor (Fable 5, 80.3%). The cost claim does not survive contact with the independent measurement: Artificial Analysis measures $3.69 per completed index task against Fable 5’s $3.08–$3.25, about 20% more, because Fable 5.1 emits roughly double the median output tokens (140M against a 71M median) — cheaper tokens, dearer tasks; on the re-based v4.3 suite the gap is starker still, $7.63 per task against Astra’s $3.26 for the same score. And on Anthropic’s own GDPval-AA v2 knowledge-work numbers it posts 1853, under the 1861 Anthropic published for Opus 5, a model at half the price. It is also slow: 66 output tokens per second and a 285-second time to first token. Mythos 5.1 takes #4 on the same basis we gave Mythos 5: it is the same base model with the cyber and bio safeguards relaxed, and it leads Anthropic’s launch table on Terminal-Bench 4.0 at 60.9% against Fable 5.1’s 55.8%. The placement is deliberately narrow. Anthropic published one benchmark for it, dropped the ExploitBench and BioMysteryBench figures that made Mythos 5’s case in June, and there is no independent scoring and never will be while access runs through the US-only Cyber Verification Program and Life Sciences Verification Program. We rank it #4 and mark it Restricted, which is the honest description of a capability you cannot buy.

At #6: GPT-6 Astra (3 September 2026) is OpenAI’s generational flagship, and the evidence for it changed under our feet. We placed it at #6 on launch week’s independent reading — Artificial Analysis scored it 61 on its Intelligence Index, level with the GPT-5.6 Sol it replaces and five points under Fable 5.1 — and then Artificial Analysis re-based the index twice in four days: v4.2 on 4 September dropped the saturated GPQA Diamond and added two private knowledge-work sets, and v4.3 on 7 September swapped Terminal-Bench 2.1 for the harder 4.0 and added AutomationBench. On the new scale Astra and Fable 5.1 tie at 53, above Opus 5 (51), Fable 5 (50) and Sol (47), and Astra gets there at $3.26 per completed index task against Fable 5.1’s $7.63 and Opus 5’s $5.86, on roughly a third of Fable 5.1’s output tokens. The agentic components are where it wins: the independently run Terminal-Bench 4.0 at 59.1% against Fable 5.1’s 52.0% and Opus 5’s 49.0%, AutomationBench-AA at 68.5%, and a Coding Agent Index of 62, level with Fable 5.1 and above Opus 5’s 60. It also leads the WebDev Arena. Three things keep it at #6 rather than #5, and out of the best-overall pick. Fable 5.1 leads it where the work is not agentic — GDPval-AA v2 knowledge work 1764 to 1662, SciCode, factual accuracy on AA-Omniscience, and Humanity’s Last Exam with tools (65.0% to 57.2%, on OpenAI’s own table) — so a tie on the composite does not displace the incumbent. It still publishes no SWE-bench figure of any kind, which costs it under our coding-weighted methodology against Opus 5 and Fable 5. And the first fortnight’s independent reviews are split: strong on games, 3D, computer use and anything with a verifiable answer; weaker on writing, with an editorial-writing Elo of 1995 against Sol’s 2156 in Decrypt’s round-up; and prone to explaining a task instead of doing it and to taking liberties with custom instructions (Peter Yang, 9 September). The capability in the headlines — the first Critical cyber rating under OpenAI’s Preparedness Framework, a perfect ExploitBench score, two zero-days found unaided — remains gated to the Daybreak Blue program, so we rank the withheld-capability model that is actually on sale at $10/$50, now on every paid ChatGPT tier and live in the API. OpenAI’s Greg Brockman used the launch briefing to declare the “AGI era”; nothing on this board moves on a press line, and it did not — the move is the re-based index.

At #15: Grok 4.7 (21 September 2026) is SpaceXAI’s first new base model since Grok 4.5, released nine days after its 12 September target. Artificial Analysis scored it 46 on the Intelligence Index v4.3 at xhigh effort, one under GPT-5.6 Sol (47) and above GLM-5.3 (45) and Grok 4.6 (44), with 56 on its Coding Agent Index (up 9) and 33% on Terminal-Bench 4.0 (up from 18%). On SpaceXAI’s own table it leads Sol on five of seven shared rows, including Terminal-Bench 4.0 (38.0% to 37.3%) and CursorBench 4.0 (46.3% to 41.7%), and trails it on DeepSWE v1.1 (71.0% to 72.7%). It sits below Sol because Sol is a point higher on the independent index and publishes a SWE-bench Pro figure (64.6%, vendor) where Grok 4.7 publishes none. The price per token is unchanged at $2/$6, but Artificial Analysis measured about 81,000 output tokens per index task against Grok 4.6’s 36,000, so a completed task costs $2.73 at high effort and $3.74 at xhigh, against $1.86 for Grok 4.6 and $1.99 for Sol; the best-value slot stays with Grok 4.6.

At #21: Gemini 3.8 Flash (2 September 2026) is Google’s fourth Flash release in under four months and the week’s economic story. Artificial Analysis scored it 59 on the Intelligence Index at high reasoning at launch — three points up on 3.7 Flash, above Opus 4.8 and Grok 4.5 (56), one under the Kimi K3/GLM-5.3 pair — and 41 on the re-based v4.3 index, a point under Opus 4.8 (42), above Grok 4.5 (39) and three to four under the Kimi K3/GLM-5.3 pair (44/45), at $0.75/$3.75 introductory pricing and $1.24 per completed task. The corroboration that matters came from a competitor: OpenAI’s GPT-6 Astra launch table prints Gemini 3.8 Flash at 73.8% on DeepSWE v1.1, within half a point of Astra (74.1%) and Opus 5 (73.7%) — at one-thirteenth of Astra’s price. Three caveats set the placement: no SWE-bench figure of any kind, an introductory price that doubles to $1.50/$7.50 on 1 January 2027, and a three-to-four-point gap to the Kimi K3/GLM-5.3 pair above it on the re-based index. Its Cyber variant is gated behind Google’s new Fairwind Program — the same week OpenAI gated Astra’s cyber capability behind Daybreak Blue.

At #24: DeepSeek V4.1-Flash (10 September 2026) is DeepSeek’s first model on a new encoder-decoder architecture, a 552-billion-parameter open-weight MoE under the MIT licence that replaced V4-Flash on DeepSeek’s API and, since 14 September, also answers requests sent to V4-Pro (DeepSeek). Artificial Analysis scores it 39 on the Intelligence Index v4.3.2 at max effort, level with GPT-5.5 and Grok 4.5, at $0.27 per completed task against $2.63 and $1.04, on $0.15/$0.60 per million tokens off-peak (double at peak). That cost per task, a tenth of GPT-5.5’s for the same score, is why it sits directly above GPT-5.5. On DeepSeek’s own table it leads Claude Opus 5 and GPT-5.6 Sol on Terminal-Bench 2.1 (90.6) and DeepSWE v1.1 (74.2). Three things keep it at #24: it is two points under Gemini 3.8 Flash (41) on the independent index, DeepSeek published no SWE-bench figure, and Artificial Analysis measured it as verbose, at 250 million output tokens to run the index against a median of 140 million.

At #14: GPT-5.6 Sol reached general availability on 9 July 2026 (ChatGPT, Codex and the API) and edges past Opus 4.8 on aggregate standing. It scores 47 on the re-based Artificial Analysis Intelligence Index v4.3, to Opus 4.8’s 42 (59 to 56 at launch), led agentic coding at launch (Terminal-Bench 2.1 and the AA Coding Agent Index) and hits 92.5% on ARC-AGI-2 — and it is token-efficient ($1.99 per index task on v4.3, against $3.26 for GPT-6 Astra and more for the Claude flagships), so on capability-per-dollar it just pips Opus 4.8 here. Two honest caveats keep the gap narrow, which is why the two sit side by side: Opus 4.8 still leads repository-scale SWE-bench Pro (69.2% vs Sol’s 64.6%, with Fable 5 ahead at 80%) and is marginally cheaper per output token ($25 vs $30), and METR flagged Sol’s reward-hacking (“cheating”) rate as the highest of any public model it has evaluated, with OpenAI’s own system card admitting task cheating — so for reliability-critical autonomous work, Opus 4.8 (just below) remains our safer “best to actually deploy” pick.

At #16: Grok 4.6 (12 August 2026) is the highest an xAI model has ever placed here, and the placement rests on independent evidence from day one: Artificial Analysis scored it 61 on the Intelligence Index at launch — tied with GPT-5.6 Sol Max for 3rd, behind only Claude Opus 5 (max, 63) and Fable 5 (max, 62), and 44 on the re-based v4.3 index (three under Sol’s 47, a point under GLM-5.3’s 45, level with Kimi K3) — with a GDPVal-AA v2 knowledge-work Elo of 1753 that trails only Opus 5, Terminal-Bench 2.1 in line with the leaders (88.4%), and $1.86 per completed task on the re-based suite ($0.84 on the old one) — under Sol, Kimi K3 and GLM-5.3 — on ~53 agent turns per task where Opus 5 uses ~103. A post-training rebuild of Grok 4.5’s foundation aimed at long-running agents, it beats its five-week-old predecessor on all nine launch benchmarks and holds the same $2/$6 price. Two things keep it below Sol under our coding-weighted methodology: it publishes no SWE-bench figure of any kind (Grok 4.5 at least reported 64.7% SWE-Bench Pro; Grok 4.6 omits the family entirely), and it loses the two hardest agentic-coding evals in SpaceXAI’s own table — DeepSWE v1.1 (65.9% vs Sol’s 73%) and Terminal-Bench 3.0 (26% vs ~34% for both Sol and Fable). SpaceXAI also publishes thinner safety documentation than its peers. But the composite is independently confirmed, the price is a fraction of the frontier’s, and unlike Sol it carries no METR reward-hacking flag — so it entered one slot down, as the frontier’s value play. Its successor Grok 4.7 now sits one place above it; Grok 4.6 remains the cheaper of the two per completed task ($1.86 against $2.73).

At #17: GLM-5.3 (14 August 2026) is the first Zhipu model to place outside the open tier on this board, on an independently measured score rather than on its licence: its weights were withheld at launch and published on 28 August under a custom licence rather than MIT. GLM-5.3 reuses GLM-5.2’s base model unchanged and takes every gain from scaled-up post-training, which is enough to move Terminal-Bench 3.0 from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9 (Z.ai). The placement rests on independent evidence: Artificial Analysis scored it 60 on the Intelligence Index at max effort — one point under the Grok 4.6 / Sol cluster and level with Kimi K3 on AA’s August re-run — and 45 on the re-based v4.3 index, a point ahead of Grok 4.6 and Kimi K3 (44) and two under Sol (47), at $1.40/$4.40 per million tokens, unchanged from GLM-5.2, for $2.01 per completed task on the re-based suite ($0.68 on the old one) — now marginally dearer per task than Grok 4.6 ($1.86) and well above Gemini 3.8 Flash ($1.24). Three things keep it at #17 rather than higher. It publishes no SWE-bench figure of any kind (GLM-5.2 reported 62.1% on Pro two months earlier), which costs it under our coding-weighted methodology. It is text-only, where every model above it takes images. And its open weights came late and not under MIT: Z.ai held them back for about two weeks after the model’s vulnerability-discovery ability compounded further than the company planned for during post-training (ExploitBench 24.4% to 54.4%), a decision Axios framed as a Chinese lab holding back an open release over hacking risk, then published them on Hugging Face on 28 August under a custom GLM-5.3 Licence that requires a Z.ai security review before commercial use by model-as-a-service businesses above $10 billion in revenue. Z.ai reports the model found 2,436 vulnerabilities across 269 open-source projects, 1,097 critical or high, including a “potentially serious” flaw in Cursor — vendor-reported, with no published CVE and no independent confirmation. With the weights out, GLM-5.3 is the highest-scoring open-weight model on the board (45 against Kimi K3’s 44), and it keeps #17.

At #18: Kimi K3 (released 16 July 2026) is the highest a Chinese model has ever placed here. The placement rests on independent evidence, not vendor claims: it scored 57.1 on the Artificial Analysis Intelligence Index v4.1 — 4th of all models at the time, above Opus 4.8’s 55.7 — and passes Opus 4.8 on GDPval-AA v2 knowledge work (Elo 1668 vs 1600), at under half Opus 4.8’s cost per task ($2.00 against $4.08 on AA’s re-based suite) (Artificial Analysis). On AA’s August re-run it was rescored to 60, just below the 61-point frontier cluster, and on the re-based v4.3 index of 7 September it reads 44 at $2.00 per task — level with Grok 4.6 and a point under GLM-5.3, which costs less than half as much per token and runs more than twice as fast — so the two mid-August arrivals that moved it down two slots stay above it. Two caveats keep it below the models above and stop us placing it higher: it has no SWE-bench Pro or Verified figure at all (Moonshot published only newer, not-yet-replayable coding suites), and Artificial Analysis measured a hallucination rate of 51%, up on its predecessor. Its open weights shipped on 26 July under the Kimi K3 License — the largest open model ever, and self-hosting avoids the hosted API’s China data residency — though at 2.8T parameters (~594GB) that’s a serious-hardware deployment. As with Sol, Opus 4.8 remains the safer deploy for reliability-critical work.

At #19: Qwen3.8-Max (3 August 2026; 0902 snapshot 2 September) is Alibaba’s 2.4T-parameter MoE flagship, with multimodal input, a 1M-token context and $2/$6 pricing. It sat here provisionally at #22 on vendor numbers alone until two things we said would move it both happened. Artificial Analysis scored it, 56 on the pre-September index and 45 on the re-based v4.3.2, level with GLM-5.3 and a point above Kimi K3. And Alibaba published open weights in August: a text-only checkpoint of the 2.4T model under its own Qwen3.8-Max License, and Qwen3.8-27B under Apache 2.0 (34 on the same index). It sits below GLM-5.3 and Kimi K3 rather than above K3 because a completed index task costs $5.41 against about $2 for both, since it is verbose (190 million output tokens to run the index) and slow (39 tokens a second), and the hosted model you can buy is not open. Its vendor figures (67.7% SWE-bench Pro, 92.6 GPQA Diamond, 86.6 Terminal-Bench 2.1) remain unconfirmed by a standardized harness.

Segmented verdicts

The single-number ranking hides the fact that “best” depends on the job. Here are the decisive picks.

Best overall (available): Claude Opus 5.5

This pick has moved to Claude Opus 5.5. It leads the independent Artificial Analysis Intelligence Index at 58 — above Sonnet 5.5 (56), Fable 5.1 and GPT-6 Astra (53) and Opus 5 (51) — at $4/$20, a fifth below Opus 5, with cache reads cut 60 per cent. Its cost per completed index task ($5.98) is level with Opus 5’s ($5.86) and below Sonnet 5.5’s ($7.60) and Fable 5.1’s ($7.63). The caveat: it publishes no SWE-bench figure, so Opus 5 (79.2% Pro, 96.0% Verified) stays our coding pick below.

Why not Sonnet 5.5, at half the token price? Claude Sonnet 5.5 is two points behind on the index, less accurate on factual questions (54% to 66% on AA-Omniscience), and at max effort uses so many output tokens that a completed task costs more than on Opus 5.5 ($7.60 to $5.98); at its default high effort it scores 47. It is the better buy where its Terminal-Bench 4.0 lead matters, or where you run it at lower effort and measure the bill yourself. Mythos 5.1 and Mythos 5 are vetted-access only. Fable 5.1 and GPT-6 Astra score 53 at $10/$50; Astra is cheaper per completed task ($3.26) but publishes no SWE-bench figure, and its first fortnight of reviews reported it taking liberties with custom instructions — the wrong trait in a default.

The one reason to still choose Opus 4.8 — and it is a real one, not a hedge: Opus 5 hallucinates slightly more. Anthropic’s own System Card concedes it “hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall”, and Artificial Analysis measured its AA-Omniscience hallucination rate rising about 14 points to roughly 50%. Opus 4.8’s headline feature was the opposite behaviour — abstaining when unsure. If your application is factual-recall-heavy and a confident wrong answer costs more than a missing one, Opus 4.8 remains the better-behaved model at the same price.

Best value: GPT-6.1 Sol

GPT-6.1 Sol (29 September 2026) takes this slot on cost per completed task, which is the measure that decides a real bill. Artificial Analysis scores it 52 on the re-based Intelligence Index, one point below GPT-6 Astra and above Claude Opus 5 (51), at $2/$10 per million tokens and $0.72 per completed index task, against $1.05 for GPT-6 Sol, which held this slot for a week, $1.86 for Grok 4.6, $1.99 for GPT-5.6 Sol and $3.26 for Astra. Claude Sonnet 5.5 shares the $2/$10 token price and scores four points higher (56), but at max effort costs $7.60 per completed task, about ten times as much, so it does not take this slot.

Three caveats. It publishes no SWE-bench figure of any kind. Its hallucination rate at max effort, though down from 60 to 54 per cent, is still high. And at launch it is in ChatGPT Work, Codex and the API, not ordinary ChatGPT chat.

The cheaper alternatives. Grok 4.6 remains the cheapest frontier-cluster model on tokens alone at $2/$6, and GLM-5.3 is cheaper still at $1.40/$4.40, if a text-only, China-hosted API suits you. Below the frontier cluster, GPT-6 Luna ($0.10/$0.50, index 37) costs $0.07 per completed task, Claude Haiku 5.5 ($0.10/$0.50 for prompts up to 100,000 tokens, index 43) $0.21, and Gemini 3.8 Flash $1.24 on introductory pricing that doubles on 1 January 2027. If you want verified coding value, GPT-5.4 still tops Scale’s standardized SWE-bench Pro leaderboard at 59.1% for $2.50/$15 (Scale SEAL).

Cheapest: GPT-6 Luna

At $0.10/$0.50 per million tokens, GPT-6 Luna (22 September 2026, covered on the GPT-6 Sol page) is the cheapest model here to run: Artificial Analysis measures it at $0.07 per completed Intelligence Index task, scoring 37 on the index. Claude Haiku 5.5 (7 October 2026) matches its token price for prompts up to 100,000 tokens, charges $0.50/$2.50 above that, and scores 43 at $0.21 per completed task. The cheapest open-weight option is DeepSeek V4.1-Flash, which replaced V4-Flash on DeepSeek’s API on 10 September 2026 at $0.15/$0.60 off-peak, doubling in peak hours (01:00–04:00 and 06:00–10:00 UTC weekdays), with MIT-licensed weights and a 1M-token context (DeepSeek).

Best free: DeepSeek V4.1-Flash (and Google’s free tier)

For a genuinely free, capable model, DeepSeek V4.1-Flash is the top model on our open-weight ranking and free to self-host under the MIT licence. The DeepSeek app is free too, with no paid tier, though DeepSeek has not said whether it now runs V4.1-Flash. For a free consumer chat from a US lab, Google’s free tier runs Gemini 3.6 Flash, with Gemini 3.8 Flash reserved for Google AI Pro and Ultra subscribers (Google), and OpenAI began rolling GPT-6 Luna out to ChatGPT’s Free tier on 22 September. (China data-residency caveats apply to DeepSeek’s hosted app; self-hosting sidesteps them.)

Best open-weight: DeepSeek V4.1-Flash

DeepSeek V4.1-Flash (10 September 2026) scores 39 on the independent Artificial Analysis Intelligence Index, above DeepSeek’s larger V4-Pro, and its MIT licence makes it self-hostable outright at 552B parameters, with a 1M-token context and image input, or $0.15/$0.60 off-peak on DeepSeek’s API. DeepSeek V4 keeps the highest open SWE-bench Verified score (80.6%, vendor-reported). Close alternatives: MiniMax M3 (80.5%, 1M context), Kimi K2.6 (80.2%), GLM-5.2 (MIT, long-horizon agents) and Llama 4 (up to a 10M-token window on Scout). New in the tier: Inkling (Apache 2.0), the only open model that reasons natively over text, images and audio — its Small variant posts a vendor 80.2% on Verified. On raw capability the open crown is now GLM-5.3’s, whose weights shipped on 28 August under a custom licence: 45 on the independent index, a point above Kimi K3 (44, open since 26 July). DeepSeek V4.1-Flash (39) is the most practical open model: MIT, far cheaper and easier to run.

Best for coding: Claude Opus 5

On repository-scale software engineering, Claude Opus 5 is the pick — 79.2% SWE-bench Pro and 96.0% SWE-bench Verified on Anthropic’s harness, against Opus 4.8’s 69.2% and 88.6% at the same $5/$25 price — and it pairs with Claude Code, the developer-favourite agentic harness. The true coding ceiling is now split. On SWE-bench Pro it remains Fable 5 (80.3%, vendor), ahead of Opus 5 by a slim margin at twice the price — because its successor Fable 5.1 published no SWE-bench figure of any kind. On agentic terminal coding the public ceiling is now Claude Sonnet 5.5 (28 September): Terminal-Bench 4.0 at 70.6% on Anthropic’s harness, above Opus 5.5 (66.4%), the restricted Mythos 5.1 (60.9%), GPT-6 Astra (57.7%) and Fable 5.1 (55.8%) — and 64% on Artificial Analysis’s own run, against about 60% for Opus 5.5 and Astra, 52.0% for Fable 5.1 and 49.0% for Opus 5. Neither Sonnet 5.5 nor Opus 5.5 publishes a SWE-bench figure, and Sonnet 5.5 costs $7.60 per completed index task at max effort, so Astra (Coding Agent Index 62, $3.26 per task) remains the cheaper agentic-coding option. The budget entry is Gemini 3.8 Flash, within half a point of Astra on DeepSWE v1.1 on OpenAI’s own table at one-thirteenth the price. On a standardized harness rather than a vendor one, GPT-5.4 still tops Scale’s SWE-bench Pro leaderboard at 59.1%. Full breakdown in our best AI for coding guide.

Best for reasoning and knowledge: Claude Opus 5

Claude Opus 5 leads Humanity’s Last Exam — the hardest general-reasoning benchmark still in rotation — in both settings, at 64.7% with tools and 56.3% without, against Opus 4.8’s 57.9% and 49.8%. On economically valuable knowledge work it scores 1861 Elo on GDPval-AA v2, the top of the board and well clear of Opus 4.8’s 1600 on the same version. (Opus 4.8’s often-quoted 1,890 is the earlier GDPval-AA, on a different scale — the two versions are not comparable, so read the version before comparing.) The caveat carried from the overall pick applies here too: Opus 5 hallucinates somewhat more than Opus 4.8, so for factual-recall-heavy work the older model is still the safer behaviour. For the hardest maths, science and logic specifically, Gemini 3 Deep Think is the dedicated ultra-tier reasoning mode worth testing.

Best long-context: Claude Opus 5.5

This pick has moved to Claude Opus 5.5. It has the same 1M-token window as Claude Opus 5, which held this slot, at a fifth lower price ($4/$20), with the highest score on the independent Artificial Analysis Intelligence Index (58). The pick rests on window, price and overall capability: no independent test we cite measures retrieval across a full 1M-token window for these models. The strongest published long-context result is OpenAI’s own for GPT-6 Astra, 96.3% on MRCR v2 8-needle between 512K and 1M tokens (vendor-reported), with a 1.05M window at $10/$50. Gemini 4 Argon, generally available since 13 October, matches the 1M input window and can also write up to 1M tokens of output, at $2/$10 until 30 November and $4/$20 after. Gemini 3.5 Pro’s announced 2M-token window is still unreleased. For the single largest raw window, Llama 4 Scout reaches 10M tokens (open-weight).

Best for privacy and self-hosting: DeepSeek V4 (MIT)

For air-gapped or compliance-bound deployments, DeepSeek V4 under MIT is the standout — frontier-adjacent quality you can run on your own hardware. For EU data residency specifically, Mistral’s Large / Medium 3.5 line is the European pick; GLM-5.2 (MIT) and Llama 4 round out the self-host options. (Serving stack matters: use bf16 rather than fp8 quantisation to preserve quality.)

What changed this month

The most recent movements on the board. For the running, dated ledger of every launch, price change and incident, see the AI changelog.

Google made Gemini 4 Argon generally available (13 October) — it stays at #7 and joins the value chart. Gemini 4 Argon reached Google AI Pro and Ultra subscribers in the Gemini app and paid customers on the Gemini API, at $2/$10 until 30 November and $4/$20 after. It keeps its place at #7, below GPT-6 Astra, and its restricted label comes off.

Meta shipped Muse Wonder and Muse Wonder Full (12 October) — they enter at #9 and #3. Muse Wonder is on the Meta Model API at $4/$20 and scored 52 on the Artificial Analysis Intelligence Index, level with GPT-6.1 Sol, at $4.52 per completed task, so it enters at #9, directly below Sol. Muse Wonder Full scored 56, level with Claude Sonnet 5.5, and enters at #3, directly below it; it is available only to Muse Code subscribers, with no API, so it stays off the value chart. Mythos 5.1 to GPT-6.1 Sol move down one place and everything from Mythos 5 down moves two, with Fable 5 and Opus 5 leaving the top ten.

Anthropic shipped Claude Haiku 5.5 (7 October) — it enters at #18. Claude Haiku 5.5 costs $0.10/$0.50 per million tokens for prompts up to 100,000 tokens, 90% below Haiku 4.5, and $0.50/$2.50 above. Artificial Analysis scored it 43 at $0.21 per completed task, so it enters at #18, directly below Qwen3.8-Max, and Gemini 3.8 Flash and everything below it move down one place.

Mistral previewed Mistral Large 4 (6 October) — it enters at #27, in preview. Mistral Large 4 is a 1-trillion-parameter model on Mistral’s API at $1.36/$4.18, with open weights promised by the end of October. Artificial Analysis scored it 38, level with Claude Sonnet 5, so it enters at #27, directly below Sonnet 5, and Gemini 3.1 Pro and everything below it move down one place.

Google announced Gemini 4 Argon (30 September) — it enters at #6, restricted. Gemini 4 Argon scored 53 on the Artificial Analysis Intelligence Index, level with GPT-6 Astra, at an introductory $2/$10, with a 1M-token output limit. Access is limited to cyber defenders in Google’s Fairwind Program, with paid API customers and Google AI Ultra subscribers next and no date given, so it enters at #6 marked Restricted, directly below Astra, and stays off the value chart until it is generally available. Everything from GPT-6.1 Sol down moves one place, and GPT-6 Sol leaves the top ten.

OpenAI shipped GPT-6.1 Sol (29 September) — it enters at #6 and takes best value. GPT-6.1 Sol keeps GPT-6 Sol’s $2/$10 and scored 52 on the Artificial Analysis Intelligence Index, four points above GPT-6 Sol and one below GPT-6 Astra, at $0.72 per completed task. It enters at #6, directly below Astra; everything from Mythos 5 down moves one place, GPT-6 Sol to #10, and GPT-5.6 Sol leaves the top ten.

Anthropic shipped Claude Sonnet 5.5 (28 September) — it enters at #2. Claude Sonnet 5.5 keeps Sonnet 5’s $2/$10 and scored 56 on the Artificial Analysis Intelligence Index, second only to Opus 5.5 (58), with Terminal-Bench 4.0 at 70.6% on Anthropic’s harness and 64% on Artificial Analysis’s — so it enters at #2, above the restricted Mythos 5.1. At max effort it uses about 193,000 output tokens per index task, the most Artificial Analysis has measured, for $7.60 per completed task, and it publishes no SWE-bench figure. Everything from #2 down moves one place.

Anthropic and OpenAI both cut prices (22 September) — Opus 5.5 takes #1, GPT-6 Sol enters at #8. Claude Opus 5.5 shipped at $4/$20, a fifth below Opus 5, and scored 58 on the Artificial Analysis Intelligence Index, the highest it has measured and five clear of GPT-6 Astra and Fable 5.1 — so it takes #1, moving Mythos 5.1 and Fable 5.1 to #2 and #3. Its cost per completed task is unchanged at $5.98, because it spends 1.6 times Opus 5’s output tokens, and it publishes no SWE-bench figure. The same day OpenAI halved API prices with GPT-6 Sol ($2/$10, index 48) and GPT-6 Luna ($0.10/$0.50, index 37), billed as standard rates: Sol enters at #8 and takes the best-value slot at $1.06 per completed task, against $1.86 for Grok 4.6. Everything from #7 down moves two places.

SpaceXAI shipped Grok 4.7 (21 September). Grok 4.7 is built on a new, larger base model, at the same $2/$6 as Grok 4.6. Artificial Analysis scored it 46 on the re-based index, one under GPT-5.6 Sol, and it enters at #8, moving Grok 4.6 and everything below it down one place. It uses more than twice Grok 4.6’s output tokens per task, so Grok 4.6 keeps the best-value slot.

DeepSeek shipped V4.1-Flash (10 September) — it enters at #19. DeepSeek V4.1-Flash is an open-weight (MIT) model at $0.15/$0.60 off-peak that replaced V4-Flash on DeepSeek’s API; since 14 September it also serves requests sent to V4-Pro, which DeepSeek is phasing out. Artificial Analysis scored it 39, level with GPT-5.5 and Grok 4.5, at $0.27 per completed task, so it enters at #19, directly above GPT-5.5.

Artificial Analysis re-based its Intelligence Index (4–7 September) — GPT-6 Astra moves up to #3 (11 September). Two interim updates on the road to the index’s v5 — v4.2 on 4 September (GPQA Diamond out as saturated; AA-Briefcase and GDP.pdf in; private test sets doubled to 40% of the weighting) and v4.3 on 7 September (Terminal-Bench 2.1 replaced by the harder 4.0; τ³-Banking replaced by AutomationBench-AA) — pulled every score down and changed the order at the top. On v4.3 Claude Fable 5.1 and GPT-6 Astra tie at 53, ahead of Opus 5 (51), Fable 5 (50), Muse Spark 1.3’s preview-only max config (48) and Sol (47); Astra costs $3.26 per completed task to Fable 5.1’s $7.63 and Opus 5’s $5.86. Our launch-week placement of Astra at #6 rested on the old reading (61, level with Sol), so it moves to #3 — below Fable 5.1, which still leads it on knowledge work, scientific coding and factual accuracy, and above Mythos 5, Fable 5 and Opus 5, which each drop one slot. Every index figure on this page now carries the v4.3 value unless it is dated; old-scale numbers are kept where they explain a launch-day placement. The rest of the board’s index scores fall too — Grok 4.6 44, GLM-5.3 45, Kimi K3 44, Gemini 3.8 Flash 41 — without changing their order, and the value-frontier chart is re-plotted on the new scale, with Astra added now that its API is live.

OpenAI shipped GPT-6 Astra (3 September) — a specialist generation with its headline capability withheld. GPT-6 Astra entered at #6 (now #6 — see the re-base entry above) on launch week’s split independent evidence: 61 on the Artificial Analysis Intelligence Index — level with the Sol it replaces — but 67 on the Coding Agent Index, equalling Opus 5 and Fable 5 at less than half Fable 5’s per-task cost. Computer use jumps furthest on OpenAI’s own table (ScreenSpot-Pro 92.7% vs Sol’s 76.9%), and Terminal-Bench 4.0 (57.7%) takes the public agentic-terminal lead from Fable 5.1. The price matches Fable 5.1 at $10/$50 — 2.5x Sol, and about 75% more per completed task on AA’s measurement — knowledge work regresses (~80 Elo, GDPval-AA v2), there is no SWE-bench figure of any kind, and the first-ever Critical cyber rating under OpenAI’s Preparedness Framework (perfect ExploitBench, two zero-days found unaided) applies to capability gated behind Daybreak Blue, not the model you can buy. GPT-5.6 Sol drops one to #7 and stays on sale at $4/$20 to at least 21 November.

Google shipped Gemini 3.8 Flash and gated its Cyber sibling (2 September). Gemini 3.8 Flash enters at #11: 59 on the independent index at high reasoning — above Opus 4.8 — at $0.75/$3.75 introductory pricing ($1.50/$7.50 from 1 January 2027) and $0.58 per completed task, undercutting GLM-5.3’s $0.68 as the cheapest task-cost near the frontier. OpenAI’s own Astra launch table corroborates the coding claim, printing it at 73.8% on DeepSWE v1.1 — within half a point of Astra and Opus 5. Gemini 3.8 Flash Cyber goes to vetted defenders through the new Fairwind Program; with OpenAI’s Daybreak Blue the same week, both US frontier labs now distribute cyber capability by vetting, not price list.

Meta and Alibaba shipped point releases (2–3 September). Muse Spark 1.3 arrived on 3 September at unchanged $1.25/$4.25 pricing, claiming 75.4% on DeepSWE v1.1 on Meta’s own table — but the strongest numbers, and the 62 Artificial Analysis measured, come from a max configuration still in limited preview and safety testing, so the board credits the shipping standard model and holds Muse Spark at #25 pending an independent score for what you can buy; Zuckerberg’s promised open weights remain unversioned and undated. Alibaba’s Qwen3.8-Max 0902 snapshot (2 September) was post-trained for coding and agent work and lifted its front-end CodeArena score 22 points to 1,691, first on that leaderboard — vendor-adjacent arena evidence, with still no Artificial Analysis entry, so Qwen3.8-Max stays provisional at #20.

How we rank

Our methodology is deliberately plural, because no single source is trustworthy on its own.

We start with benchmarks, but we read them critically: we show vendor-reported and standardized numbers side by side, label which harness produced each, and treat the vendor figure as a ceiling and the standardized one as a floor. We weight memorisation-resistant evals (SWE-bench Pro over Verified) and unsaturated ones (Humanity’s Last Exam over GPQA Diamond) more heavily.

We layer in independent reviews and public arenas — third-party explainers, community testing, and head-to-head reports — to catch the gap between a launch-day number and real-world behaviour. And we weight real-world testing: how a model behaves on long agentic runs, whether it flags its own uncertainty, how it holds up outside the benchmark’s distribution.

Every figure on this page is cited, and where a number can’t be verified from a primary source we write data not available rather than guess. The ranking is re-scored monthly — the date at the top is the last full pass. This mirrors the approach across The AI Rankings, including our best AI for coding and best AI apps guides.

One limitation: many headline frontier figures are vendor-run (Anthropic’s System Card numbers, for instance, are not independently replayable), and independent indices do not always have a score for the newest models. We flag those cases inline.

Frequently asked questions

What’s the best AI model right now?

The strongest model on independent evidence is Claude Opus 5.5 (22 September 2026), which scores 58 on the Artificial Analysis Intelligence Index — the highest it has measured — at $4/$20 per million tokens, and is generally available. The highest SWE-bench Pro figure on the board is still Anthropic’s Mythos-class Fable 5 (80.3%, vendor), which returned to general availability on 1 July 2026 after its 12–30 June export-control suspension was lifted; its restricted twin Mythos 5 is cleared for approved US organizations but stays trusted-access only. For most people the pragmatic best is Opus 5.5, or its predecessor Claude Opus 5 — half Fable 5’s price, no tightened classifier, 51 on the re-based Artificial Analysis Intelligence Index (v4.3) — and ahead of its own predecessor Opus 4.8 on coding (79.2% vs 69.2% SWE-bench Pro, vendor), reasoning (64.7% vs 57.9% on Humanity’s Last Exam with tools) and knowledge work, at the identical $5/$25 price. The one exception: Opus 5 hallucinates slightly more than Opus 4.8, so for factual-recall-heavy work the older model is still better behaved. On a standardized harness, GPT-5.4 tops Scale’s SWE-bench Pro leaderboard. The best value at the frontier is GPT-6.1 Sol — 52 on the re-based index at $2/$10 per million tokens and $0.72 per completed task — with Grok 4.6 cheaper on tokens alone ($2/$6) and Z.ai’s GLM-5.3 cheaper still ($1.40/$4.40), if a text-only, China-hosted API suits you. At the very top, GPT-6 Astra costs $3.26 per completed task against Opus 5.5’s $5.98 and Fable 5.1’s $7.63, while scoring five index points lower.

What’s the best free AI model?

DeepSeek V4.1-Flash — it is free to self-host under the MIT licence and scores 39 on the independent Artificial Analysis Intelligence Index, the top model on our open-weight ranking; DeepSeek’s app is also free, with no paid tier. For a free consumer chat from a US lab, Google’s free tier runs Gemini 3.6 Flash and ChatGPT’s Free tier is getting GPT-6 Luna. If you self-host DeepSeek you also avoid the China data-residency concerns that apply to its hosted app.

What’s the cheapest model for production use?

For raw token cost, OpenAI’s GPT-6 Luna at $0.10/$0.50 per million tokens (22 September 2026, on the GPT-6 Sol page) is the floor among the models on this page, matched by Anthropic’s Claude Haiku 5.5 for prompts up to 100,000 tokens ($0.50/$2.50 above that), and DeepSeek V4.1-Flash, which replaced V4-Flash on 10 September, is the cheapest open-weight API model at $0.15/$0.60 off-peak (doubling in peak hours). The smart pattern is routing: pin ~80% of routine traffic to a cheap or open model and reserve a flagship like Opus 4.8 or GPT-5.5 for the hard 20%.

Why do different rankings disagree so much?

Mostly because of the harness, not the model. The same model family scores ~52% on Scale’s standardized SWE-bench Pro scaffold and ~69% on the vendor’s tuned one — a 17-point swing on the same benchmark. Vendor launch numbers are a ceiling; standardized leaderboards are a floor; your real-world result sits in between and depends on your retrieval, context management and tooling. That’s why we rank by consensus across sources rather than parroting one leaderboard.

Is open-source good enough yet?

Increasingly, yes. DeepSeek V4 (MIT, 80.6% SWE-bench Verified), MiniMax M3 (80.5%, 1M context) and Kimi K2.6 (80.2%) are all within half a point of Gemini 3.1 Pro on Verified, at a fraction of the cost — close enough that, paired with a strong harness, an open model handles most routine work and enables self-hosted, data-sovereign deployments. Frontier proprietary models still lead on the hardest tasks and on long-horizon agentic reliability, but the gap that used to be 40–60 points wide is now a handful — and if Kimi K3’s promised weights ship on 27 July, an open-weights model will sit 4th overall on the independent Intelligence Index for the first time.