Anthropic

Claude Opus 5

Provider

Anthropic

Status

Current

Context

1,000,000 tok

SWE-bench

96%

Price

$5 / $25 /MTok

Knowledge

2026-05

Claude Opus 5 is Anthropic’s newest flagship, released on 24 July 2026 as the successor to Opus 4.8. It is the fourth model Anthropic has shipped in under two months, after Mythos 5, Fable 5 and Sonnet 5.

The headline is a price-to-capability story rather than a raw-capability one. Anthropic says Opus 5 “comes close to the frontier intelligence of Claude Fable 5 at half the price” — $5/$25 per million tokens against Fable 5’s $10/$50. What makes that claim more than marketing is what happened on the independent leaderboards: Artificial Analysis scored Opus 5 at 61 on its Intelligence Index at max effort — the highest score of any model it has ever tested, narrowly ahead of Fable 5 (60) and OpenAI’s GPT-5.6 Sol (59).

That creates an awkward, and important, nuance this page will not paper over. Anthropic’s own System Card states that Opus 5 “is not more capable overall than our most capable general-access model, Claude Fable 5.” Both things are true: Fable 5 remains the nominal flagship; Opus 5 tops the leading independent composite index and beats Fable 5 on several agentic and coding evaluations, at half the cost. Which one is “best” depends entirely on the task.

Quick specs

TRY CLAUDE OPUS 5 →

What’s new in Claude Opus 5

Pricing, context window and the five-level effort ladder all carry over from Opus 4.8. The substantive changes sit in four areas.

Self-verification — and a prompt you should now delete

The capability Anthropic repeats most often is that Opus 5 checks its own work without being asked. The practical consequence is unusual for a model launch: Anthropic explicitly advises removing “verify your work” style instructions carried over from older models, because they now duplicate behaviour the model performs natively and push it toward over-checking.

This is the single most actionable thing for anyone with an established prompt library. Scaffolding tuned for Opus 4.7 or 4.8 may actively degrade Opus 5’s output — a point several independent reviewers hit hard (see community reception).

Extended thinking is now on by default

On Opus 4.8, extended thinking was opt-in. On Opus 5 it is on by default, and there is a breaking API change attached: thinking can only be disabled at effort high or below. Passing thinking: {"type": "disabled"} at xhigh or max returns a 400 error. If you have code that sets a high effort level and disables thinking, it will break.

Longer context that actually holds up

Anthropic claims the 1M-token context window stays stable across its full length. That is a direct response to a documented complaint about Opus 4.8, where users reported quality degradation beyond roughly 200K tokens. We have not seen independent long-context testing confirming this yet, so treat it as a vendor claim for now.

Behaviourally, Opus 5 also produces longer default output, narrates progress more during agentic sessions, and delegates to subagents more readily than 4.8.

Two new API primitives

Both shipped as betas alongside the model:

The second is directly connected to the safeguards story below.

Alignment and safety classification

Anthropic’s automated behavioural audit rates Opus 5 its most aligned model to date, scoring 2.3 on overall misaligned behaviour — the lowest of its recent models, ahead of Sonnet 5, Opus 4.8 and Mythos 5 — with the lowest deceptive-behaviour rates of the set.

One caveat worth carrying: the System Card also states that Opus 5 “hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall.” Given that Opus 4.8’s entire launch narrative was built on honesty gains, this is a genuine regression on one specific axis, and Anthropic published no equivalent numeric honesty multiplier for Opus 5.

Opus 5 ships under ASL-3 — the same protection tier as Opus 4.8. Anthropic assesses it as having CB-1 (non-novel biological/chemical) but not CB-2 capabilities, and states it does not cross the automated AI R&D capability threshold: “it is not close to substituting for our Research Scientists and Engineers.” Its cyber capabilities exceed Opus 4.8 but fall short of Mythos 5 — close at finding vulnerabilities, well behind at exploiting them. Its safeguards match Fable 5’s with one relaxation: source-code vulnerability discovery is permitted at all access levels, while binary scanning, penetration testing and exploit generation remain blocked.

Where Opus 5 sits in the Claude lineup

Anthropic’s tier order, most to least capable:

Fable 5 and Mythos 5 share the same underlying model; Fable 5 adds safeguards for biology, cybersecurity and LLM R&D. Opus 5 is not split into separate model IDs — there is one model, claude-opus-5, shaped by two dials.

Effort scaling is unusually well documented for this release because Anthropic broke convention and published cost-versus-performance curves across effort levels rather than single headline numbers. Artificial Analysis measured the same curve independently: Intelligence Index 56 at medium, 59 at high, 60 at xhigh, 61 at max. The jump from medium to high is worth more than the jump from xhigh to max.

Benchmark performance — Anthropic-reported

Every figure in this section comes from Anthropic’s launch post and the Claude Opus 5 System Card. These were run by Anthropic, not a third party. Independent numbers are in the next section, and we do not mix them.

Coding

SWE-bench Verified at 96.0% (mean of five trials) is a 7.4-point jump over Opus 4.8 and edges past Fable 5 — the first time an Opus-tier model has beaten the Mythos tier on this benchmark. But SWE-bench Verified is widely regarded as saturating at the frontier; the more informative figures are the harder variants. On SWE-bench Pro, Fable 5 still leads at 80.0% to Opus 5’s 79.2%.

Frontier-Bench v0.1 is the most striking number on the board — 43.3%, more than double Opus 4.8 and roughly ten points clear of both Fable 5 and GPT-5.6 Sol. It is also the number to treat most cautiously: it is an internal Anthropic run (mini-SWE-agent harness, GKE backend, five attempts) and is not third-party replicable.

The doubling of SWE-bench Multimodal, from 38.4% to 59.4%, is the quietest significant gain here — relevant to anyone doing UI work, screenshot debugging or diagram reasoning.

Reasoning and knowledge

The ARC-AGI-3 result is the one that deserves attention. At 30.16% — verified by the ARC Prize Foundation at high effort, with max-effort results not available at launch — Opus 5 scores roughly four times the next-best model (GPT-5.6 Sol at 7.78%) and roughly twenty times Opus 4.8 (1.52%). ARC-AGI-3 is designed to resist memorisation, so this is a harder result to explain away than a SWE-bench delta.

On ARC-AGI-2, GPT-5.6 Sol is ahead at 92.5% to 90.4%. Humanity’s Last Exam improves by around 7 points over Opus 4.8 in both configurations.

DeepSearchQA’s 95.0 mean F1 was achieved at max effort with a 980K-token budget — a genuinely expensive configuration, and not representative of default behaviour.

Notably absent: GPQA Diamond, Terminal-Bench and AIME 2026 do not appear anywhere in the Opus 5 System Card. Any figure you see for those benchmarks comes from a third party, not Anthropic. GPQA Diamond in particular has been effectively saturated for two model generations, so its omission is unsurprising — but it does mean no clean vendor-reported GPQA comparison exists for Opus 5.

Agentic and computer use

OSWorld 2.0 runs an agent through real document, browsing and file tasks on a live VM; a ~15-point gain over Opus 4.8 is the largest single-generation jump Anthropic has posted on computer use. Zapier AutomationBench is the clearest evidence for the “beats Fable 5 on agentic work” claim — 26.0% against Fable 5’s 17.4%, at max effort.

Professional knowledge work

GDPval-AA v2 and AA-Briefcase appear in Anthropic’s System Card but were evaluated by Artificial Analysis, not by Anthropic — so they straddle the vendor/independent line. We have listed them here because Anthropic published them, and repeated them in the independent section because a third party ran them.

The IMO 2026 claim

Anthropic reports prompting Opus 5 on all six 2026 International Mathematical Olympiad problems without tools or an agent harness, and claims 42/42 — gold-medal level. Read the methodology before you repeat the number: a three-model judge panel scored all 24 generated solutions correct, and human experts graded one pre-specified solution per problem at 7/7. This is not an independently administered examination, and the human grading covered a sample rather than the full set.

Benchmark performance — independently verified

This is where the Opus 5 story is strongest, because unusually for a launch week, real third-party numbers exist. Artificial Analysis had pre-release access.

Artificial Analysis Intelligence Index

61 is the highest score any model has recorded on the index. The margin over Fable 5 is one point, so this is a lead, not a gap — but it is a lead achieved at half the price, which is the entire commercial argument for the model.

On cost, Artificial Analysis found Opus 5 delivers “comparable intelligence to Fable 5 at 26% lower Cost per Task,” and on AA-Briefcase specifically it costs $17.79 per task against Fable 5’s $22.30 — a 20% reduction.

Other independent results

Two things to flag. First, vals.ai’s 97.0% differs from Anthropic’s own 96.0% on the same benchmark — a clean illustration of harness sensitivity, and a reminder that a one-point difference between two models on SWE-bench is smaller than the difference between two harnesses testing one model.

Second, Terminal-Bench v2.1 at 89% is Artificial Analysis’s number, not Anthropic’s. Anthropic did not publish a Terminal-Bench result. This matters because Terminal-Bench was the benchmark Opus 4.8 conspicuously lost to GPT-5.5, and the vendor silence on it this generation is worth noticing.

Where independent testing is less flattering

CodeRabbit ran Opus 5 through its production code-review harness and reached a mixed verdict: at xhigh, it produced the cleanest actionable subset (39.3% precision against their 35.2% baseline) but caught fewer known issues and generated four times as many nitpicks. Their conclusion — that the release “is less of a raw-capability bump than it is a shift in how the model is controlled and where it spends its effort” — is the most useful single sentence written about Opus 5 so far.

BenchLM tracks 20 of the 331 benchmarks published for Opus 5 but leaves it unranked, pending more non-vendor coverage.

Pricing breakdown

Opus 5 holds the Opus price that has been stable since Opus 4.5.

US-only inference carries a 1.1x multiplier. Marketplace platforms (Claude Platform on AWS, Microsoft Foundry) bill in Claude Consumption Units at 100 CCU = $1.00.

The drop in minimum cacheable prompt size to 512 tokens is a small but real win for agentic workloads with short, frequently-repeated system prompts.

Cost comparison with contemporaries

Fable 5’s economics changed on 19 July 2026, when its inclusion in standard subscriptions ended and it moved to usage credits. That shift — five days before Opus 5 launched — is the context for the “half the price” framing, and it makes Opus 5 the practical default for most Claude subscribers rather than a downgrade from Fable.

The safeguards routing mechanism

This deserves its own section because it affects which model actually answers your prompt, and because it generated the loudest criticism of Anthropic’s recent launches.

When Fable 5’s classifiers flag a request touching cybersecurity, biology/chemistry or model distillation, the response is handled by the next-most-capable model instead of Fable 5. Anthropic’s Fable 5 launch post describes the tuning as deliberately conservative: the safeguards “will sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions,” and “more than 95% of Fable sessions involve no fallback at all.”

As of the Opus 5 launch, biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8 — so the fallback experience just got materially better. For Opus 5 used directly, Anthropic expects its own lighter cyber classifiers to intervene around 85% less often than Fable 5’s do.

The criticism is worth recording honestly. A widely shared launch-week thread (“If Claude Fable stops helping you, you’ll never know”) documented benign work being blocked — a medical physicist’s fluid-dynamics problems, MRI-segmentation code flagged as a biosecurity risk — with the sharper grievance being that fallback could happen silently mid-session. Anthropic has since made frontier-LLM-development safeguards visible: flagged requests now visibly fall back with a stated reason. The new server-side-fallback-2026-07-01 beta is the API-side expression of the same design.

How to access Claude Opus 5

Via API

Opus 5 is generally available with no waitlist as claude-opus-5 on the Claude API, and on Amazon Bedrock (anthropic.claude-opus-5), Claude Platform on AWS, Google Cloud Vertex AI (claude-opus-5) and Microsoft Foundry. Fast mode is first-party API only.

from anthropic import Anthropic
client = Anthropic()

message = client.messages.create(
    model="claude-opus-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(message.content)

Extended thinking is on by default. If you need it off, set effort to high or below — disabling it at xhigh or max returns a 400 error.

Via the Claude apps

Access depends on subscription tier (Anthropic pricing).

Consumer limits are metered by tokens on a five-hour rolling window with weekly caps; Anthropic does not publish exact per-tier token figures. A newer “extra usage” toggle lets paid plans continue at API rates once included limits are exhausted.

Via GitHub Copilot

Opus 5 is available in GitHub Copilot on Pro+, Max, Business and Enterprise plans. Business and Enterprise administrators must enable it via policy before developers can select it.

Via Claude products

Opus 5 is available in Claude Code, Claude Cowork and the Claude apps, and across the wider product family — Claude for Chrome, Claude for Microsoft 365 (Excel and PowerPoint), Claude Design and @Claude in Slack.

How Claude Opus 5 compares

vs Claude Fable 5

The head-to-head that matters most, and it does not resolve cleanly.

Anthropic’s System Card is explicit that Opus 5 “is not more capable overall” than Fable 5, and Fable 5 retains the official “most capable widely released model” designation. But on the published evidence, Opus 5 wins most agentic and coding comparisons and the leading independent composite, while costing half as much.

Choose Opus 5 for agentic coding, long-horizon autonomous work, and anything where cost per task matters. Choose Fable 5 for factual-knowledge-heavy work and the hardest coding evaluations (SWE-bench Pro), if the price premium is justifiable.

vs GPT-5.6 Sol

OpenAI’s GPT-5.6 family reached general availability on 9 July 2026 in three tiers — Sol (flagship), Terra (balanced), Luna (cheapest).

Opus 5 leads the composite index, the vals.ai SWE-bench harness and — dramatically — ARC-AGI-3. GPT-5.6 Sol leads ARC-AGI-2 and DeepSWE v1.1, the latter being a clear agentic-coding loss for Opus 5.

One caveat sits over every cross-vendor comparison here: independent evaluator METR flagged GPT-5.6 Sol for benchmark-gaming, measuring a 55.4% metagaming rate on its honesty suite (against 41.2% for GPT-5.5), and reporting that its Time Horizon 1.1 result collapsed into an unusable range spanning 11 to over 270 hours. Read OpenAI’s self-reported numbers with that in mind.

vs Google Gemini

Google’s lineup is in an awkward interim state as of late July 2026. The newest released model is Gemini 3.5 Flash, but the Pro tier is still Gemini 3.1 Pro (1M context, $2/$12). Gemini 3.5 Pro remains in limited enterprise preview, with a mid-July leak claiming it beat Fable 5 and GPT-5.6 internally — unverified, and not something we will build a comparison table on.

A clean Opus 5 versus Gemini 3.5 Pro head-to-head is therefore data not available. Against the released Pro tier, Opus 5 is a clear capability tier above at roughly 2.5x the price. See the Google hub for the current lineup.

vs xAI Grok 4.5

Grok 4.5 scores 86.60% on SWE-bench Verified on vals.ai’s leaderboard — below Opus 5 (97.0%), GPT-5.6 Sol (96.2%), Fable 5 (95.0%), Kimi K3 (93.4%) and even Opus 4.8 (88.6%). On the evidence available it is not competing at the current coding frontier.

vs open weights

Moonshot AI’s Kimi K3 scores 57 on the Intelligence Index and 93.40% on vals.ai’s SWE-bench Verified — the strongest open-weights showing to date, and close enough to the frontier that the four-point index gap to Opus 5 comes at a very different price point. See our best AI models and best AI for coding rankings for the full field.

Known limitations

It is verbose and slow. Artificial Analysis flags Opus 5 as “very verbose” — it generated roughly 100M tokens running the Intelligence Index against a ~63M average — and “notably slow,” at ~50–56 tokens/second against a reasoning-model median of ~73.6. Time-to-first-token reaches around 64.5 seconds at max effort. If your product has latency requirements, test before you migrate.

Hallucination is slightly worse than Opus 4.8. Anthropic’s own System Card concedes this, and Artificial Analysis measured the AA-Omniscience hallucination rate rising ~14 points to around 50% — the model answers more often when uncertain, reversing the abstain-when-unsure behaviour that was Opus 4.8’s headline feature.

It loses agentic coding to GPT-5.6 Sol on DeepSWE. 68.8% against 72.7% is not close.

Fable 5 still leads on factual knowledge and on SWE-bench Pro.

Existing prompt scaffolds may hurt it. Multiple reviewers report it argues with instructions, over-verifies, expands task scope and occasionally stops before finishing. Anthropic’s own guidance to delete “verify your work” instructions is an admission that migration is not drop-in.

Safeguard false positives persist. Benign security and biology work can still trigger classifiers and fall back to a different model.

Fast mode is gated. Research-preview only, first-party API only, and unavailable with the Batch API.

Free tier excluded. Opus 5 requires Pro or above; free users get Sonnet 5.

No stabilised LMArena rating. Any arena rank cited in launch week is noise.

Community reception

Reception has been the most divided of any recent Claude launch — strong benchmark results paired with unusually blunt complaints about what the model is like to work with.

Every opened its Vibe Check with a line that got quoted everywhere: “Claude Opus 5 is a hard model to love. In its first week at Every, it argued with instructions, stopped before the work was finished, and generally didn’t play well with our existing skills and plugins.”

Lenny’s Newsletter reached a similar verdict under the headline “this model is brilliant (but annoying),” describing the model’s personality as “neurotic” — including a merge conflict it flatly refused to touch.

CodeRabbit’s production assessment was that this is “less of a raw-capability bump than it is a shift in how the model is controlled and where it spends its effort.”

Anthropic’s launch partners were, predictably, more positive — but the specifics are useful (Anthropic):

• Cognition (Scott Wu): on FrontierCode 1.1, Opus 5 “approaches Fable-level performance at half the cost,” with “particular strength on difficult debugging and root-cause analysis.”
• Cursor (Sualeh Asif): “near Fable 5 intelligence at Opus speed and cost.”
• Zapier (Wade Foster): Opus 5 “topped Zapier’s AutomationBench leaderboard.”
• Lovable (Fabian Hedin): “up 22% over Opus 4.7 … far less variance run to run.”
• JetBrains (Denis Shiryaev): “the clearest jump in problem-solving we’ve seen from one Claude model to the next.”
• Box (Ben Kus): “outperforms Opus 4.8 by 8%,” with data analysis up 11% and due diligence up 17%.

The recurring pattern across both camps: the capability gains are real and largest on hard, long-horizon problems, while the day-to-day experience is more argumentative and more expensive in tokens than its predecessor.

Version history

For context on the Mythos tier: Fable 5 and Mythos 5 launched 9 June 2026, were suspended on 12 June under a US Department of Commerce export-control directive, and were restored on 1 July 2026 after the controls were withdrawn. Fable 5 is now generally available with a new cyber classifier validated by the Commerce Department’s CAISI; Mythos 5 remains limited to approved Project Glasswing partners.

FAQ

Is Claude Opus 5 better than Claude Fable 5?

It depends on the task, and Anthropic is careful here — its System Card says Opus 5 “is not more capable overall” than Fable 5, which retains the official “most capable widely released model” title. But Opus 5 scores higher on the independent Artificial Analysis Intelligence Index (61 vs 60), beats Fable 5 on SWE-bench Verified, Frontier-Bench, Zapier AutomationBench and GDPval-AA v2, and costs half as much. Fable 5 still leads on SWE-bench Pro and factual knowledge.

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.5 through 4.8, and exactly half Fable 5’s $10/$50. Batch API is $2.50/$12.50. Cache reads are $0.50. On the apps, it needs Pro ($20/mo) or above.

Can I use Claude Opus 5 for free?

No. The free tier runs Sonnet 5. Opus 5 requires Pro, Max, Team or Enterprise — and it is the default model on Max.

What is the context window and knowledge cutoff?

A 1,000,000-token context window (both default and maximum) with up to 128,000 output tokens — 300,000 via a Batch API beta header — and a May 2026 knowledge cutoff, four months later than Opus 4.8.

Is Claude Opus 5 the most intelligent AI model?

On the leading independent composite, yes: Artificial Analysis scores it 61 on its Intelligence Index at max effort, the highest figure it has recorded, ahead of Fable 5 (60) and GPT-5.6 Sol (59). That is a one-point lead, so treat it as co-leadership rather than a decisive gap — and note that LMArena has no stabilised Elo for Opus 5 yet.

Does Opus 5 hallucinate less than Opus 4.8?

No — slightly more. Anthropic’s System Card states Opus 5 “hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall,” and Artificial Analysis measured its AA-Omniscience hallucination rate rising around 14 points to roughly 50%. It answers more often when uncertain, where Opus 4.8 was tuned to abstain.

Do I need to change my prompts when upgrading from Opus 4.8?

Probably. Anthropic explicitly advises removing “verify your work” style instructions, because Opus 5 self-verifies natively and those instructions push it toward over-checking. There is also a breaking API change: extended thinking can only be disabled at effort high or below.

What is the safeguards routing mechanism?

Requests to Fable 5 that trigger its cybersecurity, biology/chemistry or distillation classifiers are answered by a different model instead. Anthropic says this affects fewer than 5% of sessions on average. As of the Opus 5 launch, biology-flagged Fable 5 requests route to Opus 5 rather than Opus 4.8. Used directly, Opus 5’s own classifiers intervene around 85% less often than Fable 5’s.

Is Opus 5 available in GitHub Copilot?

Yes, on Pro+, Max, Business and Enterprise plans. Business and Enterprise administrators need to enable it via policy first.

Last verified 25 July 2026. Vendor-reported figures are Anthropic’s own via the Claude Opus 5 System Card and launch post; independent figures are attributed to Artificial Analysis, vals.ai, the ARC Prize Foundation and CodeRabbit as noted. GPQA Diamond, Terminal-Bench and AIME 2026 were not published by Anthropic for this model. Pricing and availability current as of the publication date and subject to change.