Meta发布Muse Wonder系列模型，对标前沿Agent能力

We have been coveringe MSL’s 14 month journey to becoming a frontier lab, but only today did the final ball drop: Meta’s Muse Wonder to match Spark and Glimmer, the way that OpenAI has Sol/Terra/Luna and Anthropic has Opus/Sonnet/Haiku. Interestingly, while Muse doesn’t claim a specific size, Muse Wonder Full does beat Fable and Astra on some benchmarks, and with various degrees of posttraining it is a little unclear if Wonder is more of a Fable/Astra class model or not.

We have included a full breakdown of launch details below.

AI News for 10/10/2026-10/12/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: Meta launches Muse Wonder and Muse Wonder Full

What happened

Meta Superintelligence Labs (MSL) shipped Muse Wonder (codename “Watermelon” 🍉), its second-generation model family. Third-party evals put it near the frontier, roughly 14 months after the lab was founded.

• The run-up. Hours before launch, insiders posted watermelon teasers: @rpnickson (”It is a BIG watermelon…”), @alexandr_wang, @finkd (”Muse ordered me 🍉”) and @EdwardSun0909. @kimmonismus relayed leaked pricing from @synthwavedd: $4/M input, $20/M output, a 1M context window, and the max-reasoning “Wonder Full” tier exclusive to MuseCode Power/Maximum Usage plans.
• Two models. @finkd announced the family. He explained that Muse Spark was MSL’s first family and Wonder is “a much larger and more advanced class of models that brings us to the frontier.” The two releases are Muse Wonder (fast frontier) and Muse Wonder Full (the full model).
• Availability is split (@AIatMeta, Newsroom):
• Wonder is available via the Meta Model API and Muse Code.
• Full is available only to Muse Code subscribers.
• Wonder will power the Muse personal-agent product, subscribers first (@finkd).
• Muse Code installs with curl dev.meta.ai/install.sh | bash (@alexandr_wang).
• Positioning. Meta pitches the models as SOTA for “personal and professional agentic work.” @alexandr_wang named three headline features:
• Active Memory: memory across days, weeks and months.
• Active Focus: working through blockers to finish tasks.
• Domain expertise: SOTA in finance and legal, “strong in health.”
• Training story. Wang says Meta continued scaling both pre-training and RL since Muse Spark shipped in April (post; research blog). @AIatMeta framed it as going “from rebuilding our foundation to pushing the frontier in 14 months.”
• Safety claims. @AIatMeta says the family was trained for robustness to adversarial instructions and for careful handling of private information. No detailed system card was referenced.
• Launch-day usage-limit bug. Wonder and Wonder Full were draining Muse Code quotas abnormally fast (@alexandr_wang, follow-up). It was fixed the same day, with a quota reset for all users (fix; @mjdouglas). Wang noted “there’s definitely still issues we’re trying to iron out” (post).
• Media. Zuckerberg and Wang went on Stripe’s Cheeky Pint (@collision). Chapters include “Betting $100B,” agentic commerce, OpenRouter, “Why AI won’t be winner-take-all,” and MSL org structure.

Independent evals (facts)

Artificial Analysis (@ArtificialAnlys) — the most detailed third-party read:

• Intelligence Index v4.3: Wonder Full (max) scores 56 and Wonder (max) scores 52, up from Muse Spark 1.3 (max) at 48. The gap between Spark 1.3 and Wonder releases was only 6 weeks. Opus 5.5 leads at 58 (@TheRundownAI).
• Wonder Full improved on all ten index evals. The largest gains were agentic:
• Terminal-Bench 4.0: 33% → 57%
• GDPval-AA v2.1: 1681 → 1857 Elo
• AA-Briefcase: 1581 → 1729 Elo
• Where Wonder Full ranks: top three on 6 of 10 evals.
• #1 on GDP.pdf (33%) and on CritPt (33%)
• #2 on GDPval
• #3 on AA-Briefcase, AutomationBench-AA (71%) and SciCode (63%)
• Wonder (non-Full) gains: Terminal-Bench 33% → 52%; AutomationBench 58% → 68%.
• Token usage per index task:
• Output: both models use about 54k, down ~10% from Spark 1.3’s 60k. For comparison: Gemini 4 Argon 62k, Fable 5.1 78k, GPT-6.1 Sol 38k, GPT-6 Astra 27k.
• Input: Full uses 6.44M versus Wonder’s 7.69M.
• Cost per index task: Wonder costs $4.52, driven by high input usage. The table also lists:
• ModelCost per taskFable 5.1$7.63Opus 5.5$5.98Sonnet 5.5$5.46Astra$3.26Argon$1.99Sol$0.72Spark 1.3$1.60
• Wonder Full has no published token price, so its cost can’t be computed.
• Coding Agents Index: Wonder Full in Muse Code scores 68, tying Claude Code + Sonnet 5.5 (68) and ahead of Opus 5.5 (66). Wonder scores 63; Spark 1.3 scored 54.
• Specs:
• 1M context
• Input: text, image, video and audio; output: text only
• Pricing: $4 input / $0.20 cached input / $20 output per M tokens

Vals AI (@ValsAI):

• Wonder Full (Max): #5 on the Vals Index, +6.4 points over its predecessor and ahead of every OpenAI model.
• ProofBench (private): 94/100, versus 58 for Spark 1.3. The top models score 99–100 (post).
• Terminal-Bench 4: more than doubled, tying Fable 5.1 for #4 (post).
• It used about 30% fewer output tokens on Terminal-Bench, 40% fewer on ProofBench, and half the steps on Vibe Code Bench for the same score (post).
• Wonder: #10, up 7.9 points (post).
• About 37 min per task versus 57 for Full.
• Code Migration 62.6% (#6).
• Terminal-Bench 4 +38.9 points with about 20% fewer tokens.

Meta-reported headline numbers (via @kimmonismus, @Yuchenj_UW, @TheRundownAI):

BenchmarkWonder FullComparisonSWE-Marathon57.5%Opus 5.5 46.3%, GPT-6 Astra 42.5%DeepSWE v1.179.0%Opus 74.2%, Astra 74.1%Harvey LAB13.1Astra 8.6MemoryBench85.5Opus 81.5

Other third-party signals:

• @cline confirms Wonder leads DeepSWE.
• @OfirPress flagged strong ProgramBench results.
• @iScienceLuvr notes HealthBench Pro is on par with Fable 5.1.
• @AravSrinivas says it compares favorably to Fable 5.1 on Perplexity’s WANDR evals.
• Arena added it to Agent Arena and Code Arena; scores are pending (@arena). @petergostev ran cost-annotated one-shot comparisons.

Independent partner notes from @RayanKrishnan:

• Token efficiency is the biggest unlock.
• Zero refusals on their cyber benchmarks, where Wonder outscored Full.
• MSL has made “serious progress on alignment.”
• The models are strong executors but weaker at scientific-discovery intuition; their RSI Index results are pending.
• META stock rose 3% on the day.

Plots Meta emphasized:

• Performance on non-post-trained tasks versus post-trained tasks, meant as evidence against benchmaxxing and as “big model smell.” @giffmana singled this out approvingly; @scaling01 confirmed the ID-vs-OOD plot is Meta’s.
• Out-of-distribution gains over Spark 1.3, highlighted by @ren_hongyu and @jack_w_rae.

Meta’s own framing (opinion/vendor claims)

• @shuchaobi gave the most substantive insider account:
• Optimization targets “personal empowerment,” not job automation.
• Major work on very long-horizon instruction following.
• Explicit training against “laziness,” so the model is persistent before giving up.
• Defining and grading behaviors, with alignment investment to counter reward hacking under high optimization pressure.
• Claims the model one-shots most IMO problems and some open research problems with scaffolding.
• He separately describes using it for immigration, house-buying and estate planning (post).
• @alexandr_wang says research “starts with evals”: MSL built an internal Personal Superintelligence Index (PSI) rather than chasing external benchmarks, and claims “big model smell (which we quantified!).”
• @_jasonwei: “scaling prevails.” Much of Meta pivoted to support the model, including business development teams hiring human experts for data review. He also calls it “great for multimodal.”
• Other MSL staff:
• @hwchung27 — “something magical about big models.”