agents

Best AI Agents

Compare the best AI agents in 2026 — ChatGPT Agent, Claude Code, Google Antigravity, Grok Build, Manus and Devin — plus the agentic models behind them (Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1) and how far to trust agent benchmarks, with picks for every use case.

Updated October 2026

Quick answer: “Best AI agent” splits three ways in September 2026. For a general agent that browses, fills forms and finishes online tasks for you, ChatGPT Agent (on every paid ChatGPT plan from $20/month, now running GPT-6 Astra) is the most capable all-rounder. For agentic coding — the most mature category — Claude Code paired with Claude Opus 5.5 is the developer favourite, with OpenAI’s Codex, SpaceXAI’s Grok Build and Google Antigravity 2.0 as the challengers. For a do-it-all autonomous “super agent” that produces finished research, slides and spreadsheets, Manus leads. The strongest agentic model you can buy is Claude Opus 5.5 (22 September), which tops the independent Artificial Analysis Intelligence Index at 58 and costs less than the Fable and Astra tier. The caveat that frames everything below: agent benchmarks are unreliable — in April 2026 researchers showed all eight major ones could be gamed to near-perfect scores without solving a single task (UC Berkeley RDI) — and Gartner expects over 40% of agentic-AI projects to be cancelled by 2027 (Gartner).

The honest answer depends on what you want the agent to do: run your computer, write your code, automate a business process, or be built into your own product. If the steps are fixed and you only want AI called inside them, a deterministic workflow tool is cheaper and easier to audit: see best AI automation tools. New to the category? Start with what is agentic AI? for the concepts, then come back here for the tools. This guide covers the whole stack — the models that give agents their reasoning, the consumer and coding agents you can use today, the enterprise platforms, and the developer frameworks for building your own — with current benchmarks, pricing and real sentiment. Two things define the 2026 moment: the tooling has genuinely crossed from demo to daily use, and the gap between the marketing and the measured reliability has never been wider.

The current state of AI agents

An AI agent is a model given tools, memory and a goal, allowed to take multiple steps on its own — browsing, running code, calling APIs, editing files — until a task is done. In 2026 that shifted from a research demo to something people use daily, but adoption and trust are moving in opposite directions.

Gartner projects that 40% of enterprise applications will embed task-specific agents by the end of 2026, up from under 5% in 2025 (Gartner). Yet only about 17% of organisations have actually deployed an agent, with 42% expecting to within a year (Gartner via Joget). The global agent market is estimated at $10.9–12.1 billion in 2026, growing 44–46% a year. The demand is real; the delivered value is patchy.

Five shifts define where the category sits now.

1. The benchmarks broke. On 12 April 2026, UC Berkeley’s Center for Responsible Decentralized Intelligence published research showing an automated scanning agent could reward-hack all eight major agent benchmarks — SWE-bench, WebArena, OSWorld, τ²-bench and others — reaching near-perfect scores without completing the underlying tasks (benchmarkingagents.com). Since then the labs have largely stopped publishing SWE-bench at all: none of Opus 5.5, Fable 5.1, GPT-6 Astra, GPT-6 Sol or Grok 4.7 shipped with a SWE-bench figure. Every headline agent number should be read with that in mind: vendor claims are a ceiling, standardised harnesses a floor, and both are gameable.

2. “Agent washing” got called out. Gartner estimates only about 130 of the thousands of vendors marketing “agentic AI” ship anything genuinely agentic; the rest are rebranded chatbots, RPA and assistants (Gartner). The same analysis forecasts 40%+ of agentic projects cancelled by 2027 on cost, unclear value and weak risk controls.

3. Agentic coding matured into a crowded race. The most reliable agents today write code. The category consolidated around Claude Code (the developer favourite), OpenAI’s Codex, and SpaceXAI’s Grok Build — now open source and running Grok 4.7 since 21 September — with Google’s Antigravity 2.0 platform and Cognition’s Devin pushing full autonomy.

4. The model layer turned over in three weeks. Anthropic shipped Claude Fable 5.1 on 1 September, OpenAI shipped GPT-6 Astra on 3 September (on every paid ChatGPT plan by 5 September), SpaceXAI shipped Grok 4.7 on 21 September, and on 22 September Anthropic released Claude Opus 5.5 and OpenAI the cheaper GPT-6 Sol. Nearly every model this page recommended in July has been superseded.

5. Agents moved into the chat window. On 16 September Anthropic merged Cowork into the main Claude app, so agent work on your own files and apps now starts from any conversation rather than a separate mode (Claude), and OpenAI says ChatGPT’s computer use runs nearly twice as fast with Astra (9to5Mac). Underneath, every major lab ships an agent SDK — Anthropic’s Claude Agent SDK, OpenAI AgentKit and Agents SDK, Google’s Agent Development Kit — alongside framework leaders LangGraph and CrewAI.

Top AI agent models

An agent is only as good as the model reasoning underneath it. With SWE-bench largely abandoned by the labs, the useful agent measures are now Terminal-Bench 4.0 (agentic command-line work), OSWorld 2.0 (operating a real computer) and two independent composites from Artificial Analysis: its Intelligence Index and its Coding Agent Index, which runs each model inside a real coding agent. The independent columns are the ones to weight.

Reading the table: Artificial Analysis figures are its re-based v4.3 Intelligence Index (7 September 2026) and its Coding Agent Index, both at maximum effort; Grok 4.7’s Coding Agent score is with Grok Build. Vendor columns come from each lab’s launch table or a rival’s table that includes the model, so treat them as a ceiling. * Fable 5.1’s OSWorld 2.0 figure uses partial-credit scoring (41.7% strict). † Sonnet 5.5’s figure is OSWorld 2.1 with partial credit, from Anthropic’s Sonnet 5.5 table, which prints Opus 5.5 at 81.8% on the same basis. n/a means no verified figure on a comparable basis, not that the capability is missing. Not listed: Claude Mythos 5.1 (60.9% on Terminal-Bench 4.0) is restricted to vetted US organisations, so you cannot buy it. For the full cross-field ranking see best AI models.

What the numbers say — and don’t

Opus 5.5 is the new default for agent builders. It leads the independent Intelligence Index, posts the best vendor OSWorld figure on the table and the second-best Terminal-Bench 4.0 figure, and costs $4/$20 — less than half the Fable 5.1 and Astra rate. The catch is token use: Artificial Analysis measured about 119,000 output tokens per index task against Opus 5’s 73,000, so the cost per task (about $5.98) is roughly unchanged. On Artificial Analysis’s own Terminal-Bench 4.0 run it scores 59.6%, level with GPT-6 Astra and well below Anthropic’s 66.4%.

Sonnet 5.5 is the strongest terminal agent, at a price. Claude Sonnet 5.5 (28 September) scores 56 on the independent index, two under Opus 5.5, and posts the best Terminal-Bench 4.0 figures on both harnesses: 70.6% on Anthropic’s and 64% on Artificial Analysis’s, against about 60% for Opus 5.5 and Astra. It costs $2/$10 per million tokens, but at max effort it used 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. Measure the bill at the effort level you will run.

Fable 5.1 and GPT-6 Astra are level at the top of coding agents. Both score 62 on the Coding Agent Index, the highest independent coding-agent reading on this table, and both cost $10/$50. Astra does it at about 60% of Fable 5.1’s cost per task, and leads on computer use and browsing in OpenAI’s own table; Fable 5.1 is Anthropic’s pick for long-horizon agentic research.

GPT-6.1 Sol is the value pick among full-size models. Released on 29 September at $2/$10, it scores 60 on the Coding Agent Index at max effort, two points below GPT-6 Astra, at $0.72 per completed index task against Astra’s $3.26 (Artificial Analysis).

Computer-use agents are past the human baseline on vendor numbers — with caveats. The original OSWorld human baseline was 72–84% depending on task (decodethefuture); Opus 5.5’s 81.8% on OSWorld 2.0 sits at the top of that band on Anthropic’s own run. Treat “human-level computer use” as true for routine flows and false for edge cases, and note how far strict scoring pulls the numbers down (Fable 5.1: 77.9% partial, 41.7% strict).

The open and low-cost tier is closing. Kimi K3 ships open weights for self-hosting; GLM-5.3 (14 August) is Zhipu’s sharpest long-horizon agent model, with open weights published on 28 August under a custom GLM-5.3 Licence; Gemini 3.8 Flash is aimed squarely at agentic work at a Flash price; and MiniMax M3 remains the cheapest capable open-weight option at $0.30/$1.20.

Why you should distrust every agent benchmark

The scores above are useful signposts, not guarantees. Harness choice alone moves results — Anthropic’s 66.4% and Artificial Analysis’s 59.6% for Opus 5.5 on the same benchmark is a seven-point gap — and the UC Berkeley reward-hacking finding (April 2026) showed all eight leading agent benchmarks can be gamed to near-perfect scores without doing the task (benchmarkingagents.com). The practical rule: weight a model’s behaviour on your workflow over any leaderboard, run a small private eval before you commit, and treat a two-point benchmark gap as noise.

Best AI agent tools and platforms compared

We’ve grouped the field by what you actually want the agent to do, and ordered within each group by current standing.

1. ChatGPT Agent — the best general-purpose agent

Price: Included with ChatGPT Plus ($20/month) and Pro ($200/month); not on the free tier Runs on: GPT-6 Astra on paid plans since 5 September 2026, inside the ChatGPT app and web Best for: Everyday online tasks — research, bookings, forms, spreadsheets

ChatGPT Agent is the most-used general agent because it’s built into the most-used app. Invoke it with /agent and it reasons, browses websites, works with your uploaded files, connects to email and document sources, fills out forms and edits spreadsheets — pausing for confirmation on anything consequential (OpenAI). With Astra, OpenAI says computer use runs nearly twice as fast as before (9to5Mac). It’s the safe default for a non-technical user who wants a computer-using assistant without wiring anything up.

Limitations: Gated to paid tiers; can be slow on long tasks; and like all browser agents it still stumbles on unusual site layouts. Best for delegated web work, not for autonomous coding.

2. Claude Code and the Claude app (+ Claude Agent SDK) — best for agentic coding and builders

Price: Included with Claude Pro ($20/month) and Max ($100–200/month); API/credit billing for programmatic use Runs on: Claude Opus 5.5, Sonnet 5.5, Sonnet 5, Haiku 5.5; Fable 5.1 included on Max and via usage credits on Pro (Claude Help) Best for: Agentic software work; building custom production agents

Claude Code is the developer-favourite agentic coding tool (46% “most-loved” in JetBrains’ 2026 survey) and pairs the best available coding model with a strong native harness — see our best AI for coding guide for the full picture. Underneath it, the Claude Agent SDK exposes the same primitives — file editing, bash execution, web search/fetch, subagents, persistent sessions, MCP support and human-in-the-loop checkpoints — so you can build production agents on the harness that powers Claude Code (Anthropic). In 2026 it became the default way developers wire Claude into long-running workflows; a separate credit pool for programmatic use went live on 15 June 2026. Outside coding, the Claude app itself became an agent on 16 September 2026, when Anthropic merged Cowork into chat: multi-step work on your own files and apps now starts from any conversation on paid plans (Claude). On 17 September Anthropic began a beta of redesigned Projects for select Pro and Max subscribers, in which Claude coordinates parallel cloud threads that share one project memory.

Limitations: Model choice is Claude-only. Heavy Opus use adds up — Opus 5.5 spends more tokens per task than Opus 5 — so route routine steps to Sonnet 5.5 or Haiku 5.5 — measuring Sonnet 5.5’s bill at the effort you run, since at max effort it uses even more tokens per task than Opus 5.5.

3. Google Antigravity 2.0 — best multi-agent platform

Price: Free tier; Pro $20/month; Ultra $100/month Runs on: Gemini 3.x, with multi-agent orchestration Best for: Running and observing several agents in parallel; Gemini-ecosystem teams

Google Antigravity 2.0 launched at Google I/O on 19 May 2026 as a standalone, agent-first platform — desktop app, CLI, SDK and a managed agent service in one release (TechCrunch). Its Manager Surface lets you spawn and watch multiple agents at once, and a Browser Subagent tests in a real browser. It also absorbed the retired Project Mariner (shut down 4 May 2026), folding Google’s computer-use lineage into the platform, and Managed Agents in the Gemini API now spin up a tool-using agent in an isolated Linux environment with one call (Google).

Limitations: Gemini-only models; early reviews flag quota limits on cheaper tiers and the usual churn of a fast-moving platform.

4. Grok Build — SpaceXAI’s open-source coding agent

Price: Included with SuperGrok ($30/month); the harness is open source Runs on: Grok 4.7 since 21 September 2026 Best for: Terminal-native coding for developers in the Grok ecosystem

Grok Build is SpaceXAI’s entry into the coding-agent race — a full-screen terminal agent that edits files, runs shell commands, searches the web and manages long-running tasks, with its harness published as open source. Since 21 September it runs Grok 4.7, which Artificial Analysis scores 56 on its Coding Agent Index inside Grok Build, up nine points on Grok 4.6. It runs a plan mode (approve, comment on or rewrite the plan before execution) and delegates larger jobs to up to eight parallel subagents, each working a plan-search-build loop (Engadget). It makes the coding-agent field a genuine contest with Claude Code and Codex, though it still trails the Anthropic and OpenAI leaders on independent coding-agent scores.

Limitations: Smaller community and ecosystem than Claude Code or Codex; Grok 4.7 published no SWE-bench figure and uses more tokens per task than its predecessor; the hosted service is tied to the Grok subscription stack.

5. Manus — best autonomous “super agent”

Price: Free (300 daily credits, Manus 1.6 Lite); Standard $20/month (4,000 credits); Customizable $40/month (8,000); Extended $200/month (40,000); Team from $20/seat Runs on: Orchestrates frontier models from other labs Best for: End-to-end deliverables — research reports, slide decks, data tasks

Manus is the leading do-it-all autonomous agent app: describe an outcome and it plans, browses, runs code and returns a finished deliverable. It orchestrates other labs’ frontier models rather than shipping its own, and prices by credits — every browse, code run or file analysis consumes them, and monthly credits don’t roll over (No Code MBA). After Chinese regulators blocked Meta’s attempted $2bn acquisition, Meta cut ties in June and Manus resumed independent operations on 1 September 2026 (TechCrunch).

Limitations: Credit costs are hard to predict on complex jobs; because it orchestrates external models, output quality tracks whatever it routes to.

6. Devin — most autonomous software engineer

Price: Core pay-as-you-go from $20 (compute units ~$2.25 each); Team $500/month (250 units at ~$2.00); Enterprise custom Runs on: Cognition’s own stack in a sandboxed cloud environment Best for: High-volume, well-scoped engineering tasks delegated like tickets

Devin, from Cognition, is the most fully autonomous coding agent: it reads tickets from Linear, Jira or Slack, plans, writes the implementation, runs tests and opens a PR — the whole loop unattended. A February 2026 update added parallel sessions and better long-task context retention. Nubank reported 8–12x efficiency gains using it on a 6-million-line migration (Idlen). Note the separate Devin Desktop — the IDE relaunched from Windsurf — covered in our coding guide.

Limitations: Strong on bounded, repetitive work; weak and expensive on novel architecture or ambiguous requirements — independent 2026 reviews land it around 7.5/10 with frequent human intervention needed.

7. Genspark — best for finished multi-step deliverables

Price: Free tier; paid plans (credit-based) Runs on: Orchestrates 30+ models Best for: “Super agent” outputs — slides, research, even automated phone calls

Genspark is a reported $1.25bn “super agent” that orchestrates 30+ models to produce done-for-you deliverables — formatted slide decks, research reports, and outbound phone calls. It sits alongside assistants like ChatGPT rather than replacing them, and competes most directly with Manus for the autonomous-deliverables use case.

Limitations: Like Manus, output depends on routing and credits can climb; narrower general-chat ability than the first-party assistants.

8. Perplexity Computer — best multi-model orchestration

Price: On Perplexity’s paid tiers (Pro $20/month; Max $200/month) Runs on: Coordinates 19+ specialised models Best for: Long-running research and workflow tasks that span many models

Launched in February 2026, Perplexity Computer coordinates 19+ specialised models to execute long-running workflows, routing each subtask to the best-suited model (Salesforce roundup). It extends Perplexity’s sourced-answer strength from single questions to multi-step jobs, and is a natural fit for anyone who already lives in Perplexity for research.

Limitations: Best on research-shaped work; not a general desktop-automation or coding agent.

9. Salesforce Agentforce — enterprise leader

Price: Usage-based enterprise pricing (contact sales) Runs on: Atlas Reasoning Engine over your Salesforce data Best for: Customer-facing and internal agents grounded in CRM data

Salesforce Agentforce is the most-deployed enterprise agent platform, with a reported 18,500+ deals across 12,500+ companies in 39 countries. Its Atlas Reasoning Engine grounds agents in your real-time CRM metadata and business logic, with a trust layer for data masking and retention controls. For companies already on Salesforce, it’s the shortest path to production agents that act on customer data.

Limitations: Most valuable inside the Salesforce ecosystem; pricing and governance need careful modelling for high-volume use.

10. Microsoft Copilot Studio — best low-code enterprise builder

Price: Consumption-based within Microsoft 365 Runs on: OpenAI and Microsoft models over Microsoft Graph Best for: Building agents inside Teams, SharePoint and Dynamics without code

Microsoft Copilot Studio is the low-code path for the ~90% of large enterprises already on Microsoft 365. It builds, customises and deploys agents grounded in your organisation’s own data through Microsoft Graph — email, documents, Teams — and embeds them across the Microsoft stack. It’s the enterprise default where governance and existing tenancy matter more than raw model choice.

Limitations: Real value is concentrated inside the Microsoft ecosystem; consumption pricing needs forecasting for heavy workloads.

Feature comparison: the full matrix

AI agent frameworks for developers

If you’re building your own agent rather than using one, the choice is a framework or SDK — covered in full, ranked and compared, in our best AI agent frameworks guide. The field settled around a handful in 2026 (JetBrains, Morph):

• Claude Agent SDK — production agents on the Claude Code harness (Python/TypeScript): tool loop, subagents, sessions, MCP, human-in-the-loop. Now sits under Anthropic Managed Agents, which add scheduling and rubric-based outcome grading.
• OpenAI AgentKit + Agents SDK — AgentKit is the build/deploy/optimise toolkit; the Agents SDK (successor to Swarm) is the code-first path for long-horizon, sandboxed agents. Note: OpenAI’s visual Agent Builder is being wound down, with discontinuation from 30 November 2026 — build in the SDK for longevity.
• LangGraph / LangChain — the most flexible orchestration layer, with deep control over graphs, state and tool routing; the default for complex custom workflows.
• CrewAI — role-based “crews” of cooperating agents; popular for multi-agent task decomposition.
• Google Agent Development Kit (ADK) and Microsoft Agent Framework — the first-party paths for teams standardising on Gemini or Azure/Microsoft stacks respectively.
• n8n — the no-/low-code option for wiring agents into automation pipelines without writing a framework yourself.

The interoperability layer matters as much as the framework: Model Context Protocol (MCP) for tools and the open Agent Client Protocol (ACP) for running one vendor’s agent inside another’s tool are becoming the connective tissue, and are a hedge against being locked to one model provider.

Use-case specific recommendations

For everyday tasks and non-technical users

Winner: ChatGPT Agent (from $20/month)

The most capable general agent in the app most people already use — it browses, books, fills forms and edits files with confirmation checkpoints. Alternatives: Gemini’s agent if you live in Google Workspace; Manus if you want a finished report or deck rather than a chat.

For agentic coding

Winner: Claude Code + Opus 5.5 (from $20/month)

The developer favourite paired with the top-scoring model you can buy. Alternatives: OpenAI Codex with GPT-6 Astra, level with Fable 5.1 on the independent Coding Agent Index; Grok Build for terminal-native Grok users; Antigravity 2.0 for parallel multi-agent work, and Devin for delegating well-scoped tickets. Full breakdown in our best AI for coding guide.

For autonomous research and deliverables

Winner: Manus, with Genspark close behind

Both take a goal and return a finished artefact — a research report, a slide deck, a populated spreadsheet — by orchestrating multiple models. Manus leads on general task breadth; Genspark is strong on formatted outputs and even automated calls. For sourced research specifically, Perplexity Computer is the pick.

For enterprise process automation

Winner: Salesforce Agentforce (CRM) or Microsoft Copilot Studio (Microsoft 365)

Choose by the system your company already runs on: Agentforce for Salesforce-grounded customer and internal agents; Copilot Studio for low-code agents over Microsoft Graph. Both ground the agent in your own data, which is what actually drives return. See our best AI for business guide.

For developers building custom agents

Winner: Claude Agent SDK or OpenAI Agents SDK, with LangGraph for complex orchestration

Use a first-party SDK when you’re standardising on one model family and want the shortest path to production; reach for LangGraph or CrewAI when you need fine-grained control over multi-agent graphs. Build on MCP so your tools stay portable across models.

For computer-use and desktop automation

Winner: Claude Opus 5.5, with GPT-6 Astra close

For agents that must operate real applications, pick a model near the top of OSWorld 2.0 — Opus 5.5 at 81.8% and GPT-6 Astra at 72.6%, both vendor-reported — and expect reliability on routine flows but supervision on edge cases.

For value and self-hosting

Winner: Kimi K3 or MiniMax M3 behind an open framework; GPT-6 Sol if you don’t need to self-host

For data-sovereign or cost-sensitive agents, run an open-weight model built for long-horizon tool use — Kimi K3, MiniMax M3 or GLM-5.2 (MIT) — inside LangGraph or another open framework. You trade ceiling for control and a fraction of the cost. If hosted is fine, GPT-6 Sol ($2/$10) has the lowest cost per completed task among frontier-adjacent models.

For the raw capability ceiling

Winner: Claude Opus 5.5, with Fable 5.1 and GPT-6 Astra for long-horizon work

Opus 5.5 is the highest-scoring model you can buy on the independent Intelligence Index. Fable 5.1 and GPT-6 Astra share the top independent Coding Agent Index score (62) at $10/$50 — reach for them on the longest autonomous runs. The restricted Mythos 5.1 sits above all three but is not on sale.

Pricing comparison: what you’ll actually pay

Agent tools (typical monthly cost)

Models behind agents (per million tokens, USD)

Cost strategy: agents multiply token use — every step, tool call and retry bills. Pin routine steps to a cheap or open model, reserve a flagship for the hard reasoning, and cap agent runs with step limits and human checkpoints. Credit-based tools (Manus, Devin) can surprise you on complex jobs, so test on a small task before scaling. Compare models on cost per completed task, not the per-token rate: Opus 5.5 is cheaper per token than Opus 5 but uses more tokens, so a task costs about the same.

What users and developers actually think

The reliability gap is the headline

Adoption plans are aggressive, but delivered value lags. Only ~17% of organisations have deployed agents, and Gartner expects 40%+ of agentic projects to be cancelled by 2027 on cost, unclear value and weak controls (Gartner). The winners are teams that aim an agent at a narrow, measurable process rather than “automate everything”.

”Agent” is an overloaded word

With only ~130 of thousands of “agentic” vendors shipping anything genuinely autonomous, buyers have learned to test the claim. The reliable end of the market is narrow and concrete — coding agents, CRM/service agents, research agents — while the “does anything” super-agent category is more impressive in demos than in unattended production.

Benchmarks are under scrutiny

The UC Berkeley reward-hacking result (April 2026) hardened a widespread developer instinct: don’t trust agent leaderboards. Practitioners increasingly run small private evals on their own tasks and weight real behaviour — does the agent recover from errors, does it know when to stop, does it flag uncertainty — over headline scores.

Coding agents are where trust is highest

The clearest ROI stories are in software. Devin’s 8–12x migration gains and Claude Code’s developer-favourite status show the pattern: agents work best where tasks are well-specified, verifiable (tests pass or they don’t), and bounded. Independent reviews still land even the best autonomous engineers around 7.5/10, with human oversight assumed.

Recent launches reshaping the market

Claude Opus 5.5 and GPT-6 Sol (22 Sep). Anthropic’s Opus 5.5 took the top independent Intelligence Index score (58) at $4/$20; the same day OpenAI released GPT-6 Sol and Luna, a cost move that halves Sol’s cost per task (Artificial Analysis).

Grok 4.7 lands in Grok Build (21 Sep). SpaceXAI’s new base model shipped in Grok Build, Cursor and the API at the same $2/$6, lifting Grok Build’s independent Coding Agent Index score by nine points.

Claude merges Cowork into chat (16 Sep). Agent work on your own files and apps now starts from any Claude conversation, with Claude Docs and Slides launching alongside (Claude).

GPT-6 Astra (3 Sep). OpenAI’s generational flagship reached every paid ChatGPT plan by 5 September, tied Fable 5.1 on the independent Intelligence and Coding Agent indices, and sped up ChatGPT’s computer use. Its full cyber capability is gated to vetted organisations.

Claude Fable 5.1 and Mythos 5.1 (1 Sep). Fable 5.1 replaced Fable 5 on paid Claude plans and the API; Mythos 5.1 stays restricted to vetted US organisations.

Manus goes independent again (1 Sep). After the blocked Meta deal, Manus resumed independent operations and is reported to be raising $500 million at a $4 billion valuation (TechCrunch).

GLM-5.3 (14 Aug). Zhipu’s long-horizon agent model shipped on its API and coding plan with open weights withheld; Z.ai published them on 28 August under a custom GLM-5.3 Licence.

Claude Opus 5 (24 Jul) and GPT-5.6 Sol GA (9 Jul). The July flagships that replaced Opus 4.8 and GPT-5.5 as the everyday agent defaults, both since overtaken.

Grok Build enters the coding-agent race. xAI shipped a terminal-native coding CLI (grok-build-0.1, 256K context, up to eight parallel subagents) to take on Claude Code and Codex (CIO Dive).

Fable 5 restored to general availability (1 Jul). Anthropic’s frontier agentic model returned after its 12–30 June export-control suspension was lifted; Mythos 5 is cleared for approved US organisations but stays trusted-access only (Anthropic).

GPT-5.6 Sol previews (26 Jun). OpenAI’s then most capable model entered a government-coordinated limited preview before reaching general availability on 9 July.

Claude Agent SDK credits go live (15 Jun). Anthropic split programmatic agent usage onto a separate credit pool billed at API rates, formalising the SDK as a first-class product (Morph).

Project Mariner shut down (4 May), folded into Gemini. Google retired its standalone browser agent and moved its computer-use capabilities into Gemini Agent and Antigravity 2.0 (Google).

Antigravity 2.0 launches at I/O (19 May). A standalone agent-first platform — desktop, CLI, SDK and managed execution — with Managed Agents in the Gemini API spinning up a tool-using agent in one call.

OpenAI AgentKit ships; Agent Builder to wind down. The build/deploy toolkit launched, but OpenAI is discontinuing the visual Agent Builder from 30 November 2026, steering developers to the code-first Agents SDK.

Agent benchmarks shown to be gameable (12 Apr). UC Berkeley’s RDI demonstrated reward-hacking across all eight major agent benchmarks — the reliability wake-up call of the year (benchmarkingagents.com).

Frequently asked questions

What’s the best AI agent right now?

It depends on the job. For general online tasks, ChatGPT Agent (in ChatGPT Plus, $20/month) is the most capable all-rounder. For coding, Claude Code with Opus 5.5 is the developer favourite. For autonomous deliverables like reports and decks, Manus leads. For enterprise process automation, Salesforce Agentforce or Microsoft Copilot Studio, chosen by the software you already run. There is no single winner — match the agent to the task.

What is an AI agent, and how is it different from a chatbot?

A chatbot answers; an agent acts. An AI agent is a model given tools (a browser, a code sandbox, APIs), memory and a goal, and allowed to take multiple steps on its own until the task is done — browsing sites, running code, editing files, calling services. A chatbot responds to one prompt at a time; an agent plans and executes a multi-step workflow with limited supervision. For the full plain-English explainer — the agent loop, the MCP and A2A standards, and the spectrum from copilot to autonomous agent — see what is agentic AI?.

What’s the best AI agent for coding?

Claude Code paired with Claude Opus 5.5 — the most-loved tool plus the top-scoring model you can buy. Close alternatives are OpenAI Codex with GPT-6 Astra, SpaceXAI’s Grok Build, Google Antigravity 2.0 for parallel agents, and Devin for fully unattended, ticket-scoped work. See our best AI for coding guide for the full comparison.

What’s the best free AI agent?

Manus has the most useful free tier among autonomous agents — 300 daily credits on Manus 1.6 Lite. Google Antigravity 2.0 has a free tier for building and running agents, and Gemini’s agent features are available on its generous free plan. For self-hosting at zero licence cost, run an open-weight model like Kimi K3 or GLM-5.2 (MIT) behind an open framework.

Are AI agents reliable enough to trust unsupervised?

Not fully. Frontier agents now reach human range on routine computer-use tasks and resolve well-specified coding tickets, but they still fail on edge cases, novel problems and ambiguous goals — and Gartner expects 40%+ of agentic projects to be cancelled by 2027. Treat agents as capable but supervised: aim them at bounded, verifiable tasks, keep human checkpoints on anything consequential, and run your own eval before trusting a benchmark.

What’s the best AI agent for business?

Match it to your stack. Salesforce Agentforce is the most-deployed enterprise platform for CRM-grounded customer and internal agents; Microsoft Copilot Studio is the low-code choice for the Microsoft 365 majority. Both ground agents in your own data, which is what drives measurable return. Full detail in our best AI for business guide.

Which AI model is best for building agents?

Claude Opus 5.5 is the best all-round choice: the highest independent Intelligence Index score (58) and the best vendor OSWorld figure, at $4/$20. Claude Sonnet 5.5 leads Terminal-Bench 4.0 on both Anthropic’s harness (70.6%) and Artificial Analysis’s (64%) at $2/$10, though it costs more per completed task than Opus 5.5 at max effort. GPT-6 Astra and Claude Fable 5.1 share the top independent Coding Agent Index score (62) at $10/$50. For value, GPT-6 Sol has the lowest cost per completed task; for self-hosting, Kimi K3 and MiniMax M3 ship open weights.

Is Grok Build better than Claude Code?

Not yet, on current evidence. Grok Build is a capable, terminal-native newcomer with a strong parallel-subagent design, and Grok 4.7 lifted its independent Coding Agent Index score to 56 — but that still trails Claude Code’s models, and its ecosystem is smaller. Claude Code has the developer-favourite track record and access to Opus 5.5 and Fable 5.1. Grok Build is worth watching, especially for Grok-ecosystem developers; Claude Code is the safer pick today.

How much do AI agents cost?

Consumer and coding agents start around $20/month (ChatGPT Agent, Claude Code, Manus Standard). Autonomous engineers bill by usage — Devin’s compute units run ~$2.25 each, and heavy runs add up fast. Enterprise platforms (Agentforce, Copilot Studio) are usage- or consumption-based. Because agents multiply token and tool use, the real cost driver is how many steps you let them take — cap runs and route routine steps to cheap models.

Can AI agents replace employees?

Not wholesale. Agents now handle a growing slice of well-specified, verifiable work — migrations, ticketed bug fixes, customer-service flows, research compilation — and are reshaping roles toward specifying, reviewing and orchestrating rather than doing every step. But they struggle with novel judgement, ambiguity and accountability, trust remains low, and most autonomous deployments still assume human oversight. The near-term reality is augmentation of well-scoped tasks, not replacement of whole roles.

The future: what’s coming

Autonomy becomes the default interface. ChatGPT Agent, Antigravity, Grok Build and Devin point to a shift from prompting to delegating — you review outcomes, not keystrokes.

Better evaluation becomes the bottleneck. The reward-hacking finding makes trustworthy agent benchmarks a research priority; expect private, task-specific evals and outcome-graded harnesses to matter more than public leaderboards.

Model access turns strategic. With Mythos 5.1 restricted to vetted organisations and GPT-6 Astra’s full cyber capability gated, who’s allowed to use a model now shapes the toolchain, so interoperability layers — MCP for tools, ACP for cross-vendor agents — become a hedge.

The agent-washing shakeout arrives. As buyers demand measurable outcomes and 40% of projects get cut, the market should consolidate around the concrete, verifiable categories — coding, service, research — and away from the everything-agent hype.

Conclusion: how to choose

The agent market is real, useful and badly over-marketed at the same time. Pick by task, insist on measurable outcomes, and supervise.

• Best general agent: ChatGPT Agent — most capable all-rounder, in the app you already use.
• Best for coding: Claude Code + Opus 5.5; Codex with GPT-6 Astra is level on independent tests; watch Grok Build and Antigravity 2.0.
• Best autonomous “super agent”: Manus, with Genspark close for finished deliverables.
• Best for enterprise: Salesforce Agentforce or Microsoft Copilot Studio, by your stack.
• Best for builders: Claude Agent SDK or OpenAI Agents SDK, LangGraph for complex orchestration.
• Best underlying model: Opus 5.5 for most agents, GPT-6 Astra or Fable 5.1 for the longest runs, GPT-6 Sol for value.
• Best value / self-host: Kimi K3 or MiniMax M3 behind an open framework.

The tools genuinely work now for bounded, verifiable tasks — and genuinely don’t for the “do everything unattended” pitch. The variable that decides your result isn’t the logo on the agent; it’s whether you point it at a measurable job, keep a human in the loop, and test it on your own work before you trust a benchmark. For the models and coding tools behind these agents, see our best AI models and best AI for coding rankings.