AI Catchup

Our Top AI Model

Our top AI model is Claude Opus 5.5 at medium reasoning effort, with GPT-6 Astra at medium effort auditing its work. Opus 5.5 costs $4 per million input and $20 per million output tokens and leads 7 of the 9 benchmarks on Anthropic's launch table. For high-volume work, GPT-6.1 Sol and Claude Sonnet 5.5 both cost $2 and $10.

The Picks

  1. #1

    Claude Opus 5.5

    Best overall (run at medium effort)

    Anthropic's September 22, 2026 release, addressable as claude-opus-5-5, costs $4/M input and $20/M output tokens, 60% less than Fable 5.1, and beats Fable 5.1 on every benchmark Anthropic published. It leads 7 of the 9 rows on Anthropic's launch table, including Terminal-Bench 4.0 at 66.4% (GPT-6 Astra: 57.9%, Fable 5.1: 55.8%). Medium is its default effort and, on Anthropic's cost charts, its best value: 54.6% on FrontierCode for about $0.80 a task, above its own max-effort score, and 1576 on GDPval-AA for $0.86, above GPT-6 Astra at max effort (1542 for $4.53). Raise it to high for terminal-based coding agents. GPT-6.1 Sol costs half as much per token, but even OpenAI's own charts put Opus 5.5 ahead at max effort on AutomationBench (42.5% vs 36.1%) and Terminal-Bench Science (63.3% vs 57.0%). Know the caveats: thinking cannot be turned off, four API changes return errors on code written for Opus 5, and most cybersecurity work beyond fixing your own code reroutes to Opus 4.8.

    How to prompt Opus 5.5 and set its effort
  2. #2

    GPT-6 Astra

    Best for auditing work (run at medium effort)

    Use OpenAI's flagship as the second opinion on what Opus 5.5 produces: a reviewer from a different lab does not share Opus 5.5's training, which is the point of a second opinion. Astra has the lowest factual error rate on OpenAI's factuality chart (3.9%, against GPT-6.1 Sol's 4.1%) and a 0.5% rate on OpenAI's coding-deception evaluation, and OpenAI says it asks focused questions when the answer could change the outcome. It also leads Opus 5.5 on Terminal-Bench-Science on both labs' tables (64.6% vs 58.7% on Anthropic's, 68.1% vs 63.3% on OpenAI's), while AutomationBench splits: Astra leads on Anthropic's table (41.4% vs 40.0%) and Opus 5.5 on OpenAI's (42.5% vs 41.4%). OpenAI calls it the world's best model for computer use. At $10/M input and $50/M output tokens (gpt-6-astra), 2.5 times Opus 5.5's rate, it is too expensive to write everything, but a review reads far more than it writes, so most of its tokens bill at the input rate. Run it at medium, and raise effort when an audit needs browser testing or a long run of checks. If your review volume makes Astra too costly, GPT-6.1 Sol is the cheaper reviewer at a fifth of the token price.

    GPT-6 Astra launch coverage
  3. #3

    GPT-6.1 Sol

    Best for high-volume coding agents

    OpenAI's September 29, 2026 upgrade to GPT-6 Sol, addressable as gpt-6.1-sol, costs $2/M input and $10/M output tokens with cached input at $0.10, a fifth of Astra's token price. On OpenAI's charts it scores 75.2% on DeepSWE at high effort for about $0.65 a task, above Astra's best of 74.1% at $4.43, and it beats GPT-6 Sol's best score on all six of OpenAI's benchmark charts. Astra still leads it on OSWorld offline (73.5% vs 71.4%), AutomationBench, and Terminal-Bench Science. It is our default OpenAI model for coding agents: run it at high for coding, and put agents on the Responses API, since Chat Completions no longer supports tool calling with it.

    GPT-6.1 Sol: benchmarks and cost at every effort level
  4. #4

    Claude Sonnet 5.5

    Best value Claude model

    Anthropic's September 28, 2026 mid-tier model, addressable as claude-sonnet-5-5, keeps Sonnet 5's $2/M input and $10/M output price, and Anthropic halved its cache reads to $0.10 per million on October 7. It scores 70.6% on Terminal-Bench 4.0 against Sonnet 5's 10.3%, and 55.5% on CursorBench 4.0, within about two points of Opus 5.5's 57.8%. Anthropic still calls Opus 5.5 clearly stronger at complex, open-ended work, so give Sonnet 5.5 the well-scoped everyday tasks and bug fixes where speed and cost matter. It costs the same per token as GPT-6.1 Sol, and no benchmark is shared between the two labs' launch tables, so test both on your own tasks.

    How to prompt Sonnet 5.5
  5. #5

    Claude Haiku 5.5

    Best for subagents and high-volume calls

    Anthropic's October 7, 2026 small model, addressable as claude-haiku-5-5, costs $0.10/M input and $0.50/M output tokens for prompts up to 100,000 tokens, and five times that above. It is Anthropic's fastest model and its first Haiku with effort levels. On Anthropic's table it scores 72.4% on OSWorld 2.1 (offline subset) against GPT-6 Luna's 48.9%, and 39.2% on Terminal-Bench 4.0 against Luna's 16.4%, at the same $0.10 and $0.50 token price as Luna. Use it under Opus 5.5 or Sonnet 5.5 for subagents, summaries, compaction, and classification.

    Claude Haiku 5.5 launch coverage
  6. #6

    Grok 4.7

    Best budget pick for knowledge work

    SpaceXAI's September 21, 2026 release, addressable as grok-4.7, costs $2/M input and $6/M output tokens below 200,000 prompt tokens ($4 and $12 above), with a 500,000-token context window. Its $6 output rate is 40% below GPT-6.1 Sol's and Sonnet 5.5's $10. It is live in Cursor and Grok Build, where a fast variant runs at twice the output speed for twice the price. On SpaceXAI's own table it leads Fable 5.1 on the Harvey Legal Agent Benchmark (19.6% vs 6.7%) and EEBench (64.0% vs 56.4%). Coding is the caution row: 46.3% on CursorBench 4.0 and 37.6% on Terminal-Bench 4.0, well below Sonnet 5.5's 55.5% and 70.6%, so bake it off before making it a coding default.

    SpaceXAI's Grok 4.7 announcement
  7. #7

    Gemini 3.1 Pro

    Best for video and audio input

    Google's Pro model, still in preview as gemini-3.1-pro-preview, takes text, image, video, audio, and PDF input, with a 1,048,576-token input limit and 65,536 output tokens, at $2/M input and $12/M output tokens for prompts up to 200,000 tokens ($4 and $18 above). Pick it when the input is a recording or a video: GPT-6.1 Sol accepts text and images only. Its context window is no longer a reason on its own, since Opus 5.5 and GPT-6.1 Sol also take about a million tokens.

    Google's Gemini 3.1 Pro model page
  8. #8

    GLM 5.3

    Best open-weight

    Z.ai's latest open-weight coding model and the successor to GLM 5.2, with weights on Hugging Face under Z.ai's own GLM-5.3 license rather than a standard open-source license. Z.ai calls it the most capable open-weights model for coding, but on its own table the lead is narrow: 66.9% on DeepSWE v1.1 against Kimi K3's 67.5%. Cursor serves it with a 1M-token context window and Low, High, and Max reasoning levels. It is the option for teams that need to self-host or inspect the weights.

    GLM 5.3 in Cursor

How They Compare

EvalClaude Opus 5.5GPT-6 AstraClaude Fable 5.1
Price per 1M tokens (input / output)$4 / $20$10 / $50$10 / $50
Terminal-Bench 4.066.4%57.9%55.8%
FrontierCode v1.1 Main54.4%53.3%50.3%
GDPval-AA v2.1 (Elo)184615421735
AutomationBench40.0%41.4%31.4%
Humanity's Last Exam (with tools)67.7%57.2%65.6%
Terminal-Bench-Science 0.158.7%64.6%52.6%

Figures are from Anthropic's Claude Opus 5.5 launch table (September 22, 2026), with Opus 5.5 run at max effort unless noted. Terminal-Bench 4.0 is each model's best score (Opus 5.5 at xhigh, Astra at high), and Terminal-Bench-Science carries a standard error of 3.5 to 5 points. These are not our day-to-day settings: at medium, Opus 5.5 scores 57.6% on Terminal-Bench 4.0. Cost per task at every effort level is in our Opus 5.5 coverage.

Changelog

  • October 9, 2026: GPT-6.1 Sol replaces GPT-6 Sol at #3: same $2/$10 token price, cached input halved to $0.10, and a better best score on all six of OpenAI's benchmark charts. Claude Sonnet 5.5 replaces Sonnet 5 and moves up to #4, ahead of Grok, now Grok 4.7: Sonnet 5.5 leads it on CursorBench 4.0 (55.5% vs 46.3%) and Terminal-Bench 4.0 (70.6% vs 37.6%) at the same input price. Claude Haiku 5.5 is added at #5. Gemini 3.1 Pro moves to #7 as the pick for video and audio input, because Opus 5.5 and GPT-6.1 Sol now match its million-token window. GLM 5.3 replaces DeepSeek R1, a much older model, as the open-weight pick. Opus 5.5 and Astra keep #1 and #2: OpenAI's own charts put Opus 5.5 ahead of GPT-6.1 Sol at max effort on AutomationBench and Terminal-Bench Science, and Astra ahead of it on five of six. Corrections: the Gemini entry said multi-million-token windows, but Google lists a 1,048,576-token input limit, and its ARC-AGI-1 and BrowseComp leads came from an April table we could not re-verify, so they are removed. The Sonnet 5 entry called it the default for Free and Pro plans; Claude Code now defaults Pro to Opus 5.5.
  • September 24, 2026: Claude Opus 5.5 replaces Claude Fable 5.1 as the top pick, run at medium effort: it beats Fable 5.1 on every benchmark Anthropic published at 40% of the per-token price. GPT-6 Astra, at medium effort, becomes the #2 pick as the model that audits Opus 5.5's work, replacing GPT-5.6 Sol as the OpenAI co-pick. GPT-6 Sol replaces GPT-5.6 Sol for high-volume agent work, and Fable 5.1 rotates out. The comparison table moves to Anthropic's September 22 launch figures.
  • September 1, 2026: Claude Fable 5.1 replaces Fable 5 as the top pick on its September 1 launch, still recommended at medium reasoning effort. Same per-token price as Fable 5 with cache reads cut 75%, and Anthropic reports it ahead of Fable 5, Opus 5, and GPT-5.6 Sol on every benchmark it published. GPT-5.6 Sol stays the agent co-pick at xhigh; the comparison table remains SpaceXAI's August 12 figures, which predate Fable 5.1.
  • August 12, 2026: Claude Fable 5 promoted to the top pick, recommended at medium reasoning effort, with GPT-5.6 Sol at xhigh as the co-pick for terminal-heavy, long-horizon coding agents. Benchmarks re-based on SpaceXAI's August 12 Grok 4.6 launch table. Grok 4.6 added at #3 for price-to-capability; Claude Sonnet 5 replaces Sonnet 4.6 as the value pick; GPT-5.5, GPT-5.5 Pro, and Claude Opus 4.7 rotate out, superseded by GPT-5.6 Sol and Fable 5.
  • April 23, 2026: GPT-5.5 promoted to top overall model on its April 23 launch. State-of-the-art on Terminal-Bench 2.0 (82.7%), GDPval (84.9%), OSWorld-Verified (78.7%), FrontierMath Tier 4 (35.4%), and CyberGym (81.8%). Matches GPT-5.4 per-token latency while delivering higher intelligence and using fewer tokens on equivalent Codex tasks. Claude Opus 4.7 moves to #2 as the pick for SWE-Bench Pro-style work. GPT-5.5 Pro added as a separate tier for the hardest questions.
  • April 16, 2026: Claude Opus 4.7 promoted to top model on its April 16 launch. Notable gains on the hardest software engineering tasks, state-of-the-art on Finance Agent and GDPval-AA at that time, plus higher-resolution vision (2,576 pixels on the long edge) and a new xhigh effort level. Pricing unchanged from Opus 4.6 at $5/M input and $25/M output.
  • March 2026: Initial picks published. Claude Opus 4.6 selected as top model.

Frequently Asked Questions

Why is Claude Opus 5.5 the top pick?

It beats Claude Fable 5.1 on all nine benchmarks Anthropic published at 40% of the price, and leads GPT-6 Astra on four of the six benchmarks both are scored on, including Terminal-Bench 4.0 (66.4% vs 57.9%). GPT-6.1 Sol is half the token price, but OpenAI's own charts put Opus 5.5 ahead at max effort on AutomationBench and Terminal-Bench Science.

Why use GPT-6 Astra to audit Opus 5.5's work?

A reviewer from a different lab does not share Opus 5.5's training. Astra has the lowest factual error rate on OpenAI's factuality chart (3.9%) and leads Opus 5.5 on Terminal-Bench-Science on both labs' tables (64.6% vs 58.7% on Anthropic's, 68.1% vs 63.3% on OpenAI's). A review mostly reads, which keeps its $10 and $50 price manageable.

What reasoning effort should I set?

Medium for both. Medium is Opus 5.5's default, thinking is always on, and on Anthropic's charts medium is its best value; use high for terminal-based coding agents and xhigh or max for knowledge-work deliverables. Run Astra at medium for reviews. Our Claude Opus 5.5 prompting guide covers each effort level in detail.

Should I move from GPT-6 Sol and Sonnet 5 to the newer versions?

Yes for both. GPT-6.1 Sol costs the same per token as GPT-6 Sol, halves cached input to $0.10, and posts a better best score on all six of OpenAI's benchmark charts. Claude Sonnet 5.5 keeps Sonnet 5's $2 and $10 price and scores 70.6% on Terminal-Bench 4.0, against Sonnet 5's 10.3%.

Is Claude Fable 5.1 still worth using?

Only where your own evals show it ahead. Opus 5.5 beats it on every benchmark Anthropic published at 40% of the price, though Anthropic says the real-world gap is narrower than the scores suggest.

How often does the top model change?

We re-evaluate whenever a major model release changes the landscape. GPT-5.5 held the top slot from April 23 to August 12, 2026, when Claude Fable 5 took it; Fable 5.1 replaced it on September 1, and Claude Opus 5.5 replaced Fable 5.1 on September 24, 2026, with GPT-6 Astra as the auditing pick.

For effort settings and prompts in depth, read our Claude Opus 5.5 prompting guide.

Sources

Get the weekly AI Catchup

Tools, practices, and what matters, in your inbox every week.