How to read this page

Vendor tables report selected benchmarks and configurations; not every result is independently reproduced. Artificial Analysis applies its own published evaluation methodology, but settings and harnesses can differ by benchmark, model and evaluation version. Comparisons are strongest within the same benchmark and methodology version. Neither column is the whole story; the gap between them usually is.

Meta Muse Wonder and Muse Wonder Full (released October 12, 2026)

Meta released two models. Wonder Full posts the headline scores but is not on the API (Muse Code subscriptions only); standard Wonder is the deployable model, at $4/$20 per million input/output tokens. Full coverage: Meta Reached the Frontier. It Left the Weights Behind. — https://alfw.ai/article/meta-muse-wonder-frontier-closed/

### Artificial Analysis Intelligence Index v4.3 (10-eval composite)

Vendor-reported: Meta's model card lists the index: Wonder Full 56, Wonder 52, Opus 5.5 58 — these are AA-measured results republished by Meta. Independent: Artificial Analysis's own figures: Full 56, Wonder 52, Opus 5.5 58, Sonnet 5.5 56, GPT-6 Astra 53. Gap note: no gap here — the rare row where the vendor table and the independent table are the same table. The selection happens in which rows Meta headlines, not in this number. Sources: Meta model card via VentureBeat — https://venturebeat.com/technology/metas-new-muse-wonder-and-muse-wonder-full-models-achieve-frontier-performance-but-only-the-weaker-one-is-available-via-api ; AA v4.3 as reported by OfficeChai — https://officechai.com/ai/muse-wonder-artificial-analysis-intelligence-index/

### SWE-Marathon (long-horizon coding)

Vendor-reported: Wonder Full 57.5, Wonder 53.1, Opus 5.5 46.3, GPT-6 Astra 42.5 (Meta model card). Independent: no independently reproduced result identified as of October 12, 2026. Meta's card states this row "is not a controlled common-harness comparison" (results via Abundant AI; Wonder Full ran at an "xhigh" reasoning setting). Gap note: Meta reports Full ahead of Opus by 11.2 points (57.5 versus 46.3), but the configurations are not controlled under a common harness. The difference is a reported benchmark result, not an isolated model-performance advantage. Source: Meta model card via Implicator — https://www.implicator.ai/meta-muse-wonder-ranks-below-claude-opus/

### Terminal-Bench 4.0 (AA-run variant, agentic coding)

Vendor-reported: Meta's card publishes the AA-run row: Wonder Full 57.1, Wonder 52.0, GPT-6 Astra 59.1, Opus 5.5 59.6. Independent: Artificial Analysis's own run of Terminal-Bench 4.0 puts Wonder Full at 57%. Gap note: in the reported Terminal-Bench variant, Opus (59.6) and Astra (59.1) exceed Full (57.1). SWE-Marathon ranks them differently, but these are different tasks and evaluation configurations; the reversal cannot be attributed to the harness alone. Source: Meta model card via Maccus — https://maccus.fr/en/meta-muse-wonder/

### ProgramBench (rebuild a program from scratch)

Vendor-reported: in Meta's own table: Wonder Full 59.0, Opus 5.5 70.5. Independent: no independently reproduced result identified as of October 12, 2026. Gap note: not a vendor-vs-independent gap — a selection gap. This row sits in the vendor's own table, 11.5 points behind the rival, and appears in none of the launch headlines. Source: Meta model card via Maccus — https://maccus.fr/en/meta-muse-wonder/

### Harvey LAB-AA (legal agent work)

Vendor-reported: Wonder Full 13.1, GPT-6 Astra 8.6, Opus 5.5 4.2 (Meta model card). Independent: the benchmark is Artificial Analysis's version of Harvey's Legal Agent Benchmark; a task passes only if the deliverable meets every rubric criterion. Gap note: Meta reports Full at 13.1 on the LAB-AA evaluation, ahead of the selected rivals. The reported low absolute score is notable, but the table alone does not establish independent replication or performance across legal work generally. Source: SiliconANGLE — https://siliconangle.com/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and-professional-agents/

### Cost per Index task — Muse Wonder API, not Wonder Full (AA estimate, maximum-effort configurations)

Vendor-reported: list prices are identical for Wonder and Opus 5.5: $4 in / $20 out per million tokens. Independent: Artificial Analysis estimates an average Intelligence Index task costs $4.52 on Muse Wonder, $5.98 on Opus 5.5 and $3.26 on GPT-6 Astra. Muse Wonder Full has no published per-token price or AA cost-per-task estimate, because it is available through Muse Code rather than the Wonder API. Gap note: equal headline input/output token rates do not imply equal task cost. Total cost depends on input, output, caching and evaluation configuration; AA reports different per-task totals, and its Wonder calculation includes cache writes and cache hits. Source: AA estimates via alfw.ai coverage — https://alfw.ai/article/meta-muse-wonder-frontier-closed/

Anthropic Claude Opus 5.5 and Sonnet 5.5 (released September 22 and 28, 2026)

Opus 5.5 is Anthropic's flagship at $4/$20 per million tokens; Sonnet 5.5 is the half-price tier at $2/$10. Full coverage: Anthropic's Sonnet 5.5 Nearly Matches Opus 5.5. The Cost Math Needs a Footnote. — https://alfw.ai/article/claude-sonnet-5-5/

### Artificial Analysis Intelligence Index v4.3

Vendor-reported: Anthropic positions Opus 5.5 as its most capable model. Independent: Opus 5.5 scores 58 (1st), Sonnet 5.5 scores 56 (tied with Muse Wonder Full). Gap note: the independent aggregate confirms the flagship ordering. Note the index itself is revised over time (v4.1 to v4.3 within weeks) — scores are only comparable inside one version. Source: AA as reported by OfficeChai — https://officechai.com/ai/muse-wonder-artificial-analysis-intelligence-index/

### Terminal-Bench 4.0 (agentic coding)

Vendor-reported: Anthropic's launch table reports Sonnet 5.5 at 70.6% and Opus 5.5 at 66.4% on Terminal-Bench 4.0; the exact evaluation configurations should be checked against Anthropic's footnotes before treating them as identical. Independent: in Artificial Analysis's run of the same benchmark (the Terminal-Bench 4.0-AA variant published in Meta's model card), Opus 5.5 scores 59.6. Gap note: same benchmark name, roughly 7 points lower in the AA-run variant. The configurations differ between the two runs; the gap cannot be attributed to the harness alone — they are different measurements wearing one label. Sources: Anthropic launch figures via alfw.ai — https://alfw.ai/article/claude-sonnet-5-5/ ; Meta model card via Maccus — https://maccus.fr/en/meta-muse-wonder/

### Cost per task (Sonnet 5.5)

Vendor-reported: Anthropic advertises Sonnet 5.5 as costing up to 30% less for most work, in the context of its improvements over Sonnet 5. The comparison is not a published claim of a 30% saving against Opus 5.5. Independent: Artificial Analysis found Sonnet 5.5 burns roughly 60% more output tokens than Opus 5.5 at maximum effort (an attributed secondary-source estimate). Gap note: AA's reported maximum-effort token consumption is a separate cost comparison against Opus. It cannot establish whether Anthropic's claimed savings against Sonnet 5 hold at default or medium effort. Source: AA via alfw.ai — https://alfw.ai/article/claude-sonnet-5-5/

### Case study — ARC-AGI-3 and the harness question (Claude Opus 5)

This row is a methodological case study, not a direct replication of a vendor benchmark. Through the benchmark's official interface (a 64×64 grid of numbers), Opus 5 scores 40.68. Nvidia's AVO harness scored 100 on the public set with the same model in August 2026; MIT's VISTA harness also scored 100 (October 1, 2026) — both by giving the model screenshots and a browsable frame history instead of numbers. Gap note: the reported public-game scores range from 40.68 to 100 under different interfaces and agent harnesses. This illustrates sensitivity to the surrounding system, but does not isolate harness effects unless model version, prompts, tools, budgets and scoring conditions were otherwise controlled. Caveats stay attached: public games only, the private set has not been run, and free frame-browsing flatters the score. Source: alfw.ai — https://alfw.ai/article/vista-perfect-arc-agi-3-score/

OpenAI GPT-6 family (GPT-6 Astra, September 2026; GPT-6.1 Sol, September 29, 2026)

Astra is OpenAI's frontier tier at $10/$50 per million tokens; Sol is the $2/$10 workhorse tier. Full coverage: OpenAI DevDay 2026 — https://alfw.ai/article/openai-devday-2026-dots-sol-agents-api/

### Artificial Analysis Intelligence Index and the effort ladder

Vendor-reported: OpenAI markets Astra as its frontier model. Independent: AA places Astra at 53 on v4.3, behind Opus 5.5 (58) and Muse Wonder Full (56). AA's effort-by-effort runs (September 2026): Astra beats Opus 5.5 at low effort (46 vs 42) and trails from medium upward; at max effort, 53 vs 58. Cost per task at max effort: Astra $3.26 vs Opus $5.98. Astra emits far fewer output tokens per task (about 27k, AA-confirmed, vs Opus's about 119k, AA figures via dev.to summary). Gap note: "which model is better" is partly an effort-setting question; the token-consumption difference is why the pricier-per-token model is cheaper per task. Sources: AA data via Emergent — https://emergent.sh/learn/claude-opus-5-5-vs-gpt-6-astra ; AA token figures via dev.to — https://dev.to/genelab_999/claude-opus-55-vs-gpt-6-astra-vs-gpt-6-sol-the-effort-knob-matters-more-than-the-model-55gm

### GPT-6 Instant speed

Vendor-reported: OpenAI reports that GPT-6 Instant begins answering web-search questions 44% sooner on average than GPT-5.6 Instant, based on its internal measurement of time to first response. Independent: no independent measurement published — unknown. Gap note: vendor-reported latency improvement; the public announcement does not provide enough methodological detail for independent reproduction. No independent replication identified. It measures when the answer starts, not when it finishes. Source: OpenAI launch announcement via alfw.ai — https://alfw.ai/article/openai-gpt-6-intelligent-ui-chatgpt/

### GPT-6.1 Sol pricing framing

Vendor-reported: OpenAI describes GPT-6.1 Sol as delivering near-Astra performance at a lower price; its standard input and output token rates are one-fifth of Astra's ($2/$10 vs $10/$50). Independent check: against its own predecessor, GPT-6 Sol, the standard input and output prices did not change; the only cut was cached input, $0.20 to $0.10 per million tokens. Gap note: the "one-fifth" baseline is the expensive sibling, not the previous version of the same model. The price ratio is verified; equivalent capability is not. Source: alfw.ai analysis of OpenAI pricing — https://alfw.ai/article/openai-devday-2026-dots-sol-agents-api/

### Case study — GPT-6 Astra "blind" World of Warcraft run (illustrative demonstration, not a standardized evaluation)

Astra, inside Codex, completed every quest in WoW's Orc starting zone in about 40 minutes with zero deaths, without rendering a frame. The run's own developer notes the conditions: the agent read raw server packets, navigation meshes and the server's quest database — more direct information than a human player gets — on a private AzerothCore server, in an environment that was not fully sandboxed. Gap note: not a benchmark and not "blind" in the human sense. The tool-building is genuinely impressive; the framing converts a protocol-level run into a playing-the-game story. Source: alfw.ai — https://alfw.ai/article/gpt-6-astra-wow-blind-run/

What the gaps mean

Three patterns account for nearly every vendor-vs-independent gap on this page.

First, selection. A launch table is an edited document. Meta's card contains a row where its model loses ProgramBench by 11.5 points; that row is real, published and absent from every headline. When a vendor table and an independent index disagree about who leads, check which rows each side chose to print before arguing about the numbers.

Second, the setup around the model is part of the result. Reported agent-benchmark scores move by points — sometimes by tens of points — across interfaces, scaffolds and effort settings, even when most comparisons are not controlled tightly enough to credit any single factor. Agent benchmarks measure a model-plus-system combination. Any headline that reports the model alone is rounding off the part that often matters most.

Third, baselines and settings do quiet work. "One-fifth the price" compares a model to its expensive sibling instead of its predecessor. "Up to 30% cheaper" was claimed against the previous generation, not the flagship it gets quoted against. "44% sooner" measures time to first response and cannot be independently reproduced from the announcement. None of these statements needs to be false to mislead; each one needs its comparison class attached before it means anything.

What the independent column cannot do: Artificial Analysis's index is one methodology with its own weighting choices, revised between versions, so its scores are a consistency tool, not a verdict. The practical test remains the unglamorous one — run the deployable version of each model, at the effort setting you will actually pay for, on your own tasks, and compare cost per finished task.

Change log

October 12, 2026 — Page created with the Muse Wonder launch (Meta), the Claude 5.5 family figures (Anthropic) and the GPT-6 family figures (OpenAI) available as of this date. Artificial Analysis's October 12 Muse Wonder analysis is the primary source for its reported Muse scores; other Index scores and effort-specific comparisons are attributed to the named secondary reports unless directly confirmed against AA's underlying results. Next scheduled check: the next frontier-model launch, or any Artificial Analysis index revision.

Sources

  1. [1] Meta Muse Wonder model card, October 12, 2026 — full benchmark table, via VentureBeatRead source
  2. [2] Meta Muse Wonder model card — methodology and the "not a controlled common-harness comparison" note, via ImplicatorRead source
  3. [3] Meta Muse Wonder model card table (Terminal-Bench 4.0-AA and ProgramBench rows), via MaccusRead source
  4. [4] Artificial Analysis, Muse Wonder analysis, October 12, 2026 (primary source for AA's Muse scores); Intelligence Index v4.3 results as reported by OfficeChaiRead source
  5. [5] Artificial Analysis effort-level comparison, Claude Opus 5.5 vs GPT-6 Astra, September 2026, via EmergentRead source
  6. [6] Artificial Analysis token-consumption figures, via dev.toRead source
  7. [7] SiliconANGLE, Meta Muse Wonder launch coverage (Harvey LAB-AA scoring rules), October 12, 2026Read source
  8. [8] TechCrunch, Meta Muse Wonder launch and pricing, October 12, 2026Read source
  9. [9] alfw.ai, Meta Reached the Frontier. It Left the Weights Behind. (October 12, 2026)Read source
  10. [10] alfw.ai, Anthropic's Sonnet 5.5 Nearly Matches Opus 5.5. The Cost Math Needs a Footnote. (October 2, 2026)Read source
  11. [11] alfw.ai, OpenAI DevDay 2026 (September 29, 2026)Read source
  12. [12] alfw.ai, GPT-6 in ChatGPT / Intelligent UI coverage (October 7, 2026)Read source
  13. [13] alfw.ai, GPT-6 Astra Cleared WoW Blind (October 4, 2026)Read source
  14. [14] alfw.ai, The Perfect AGI Score Was Never About the Model (VISTA / ARC-AGI-3, October 6, 2026)Read source