Meta launched the Muse Wonder series on October 12: Muse Wonder and Muse Wonder Full. In one block:

Muse Wonder      → Meta Model API + Muse Code
Muse Wonder Full → Muse Code subscribers only, the largest model
Headline pair    → Active Memory + Active Focus

Strip away the superintelligence framing and the message is practical: coding-tool competition is moving from code generation to long-horizon execution.

Active Memory vs Active Focus

Read them as a pair:

Active Memory = what to remember
  → keep what matters from past conversations,
    pull it back when the current task needs it.

Active Focus = how to finish
  → try another approach when blocked, recover from errors,
    ask when unsure, check in before important actions.

Either half alone disappoints. Memory without focus remembers the old discussion and stalls at the first blocker. Focus without memory never quits but keeps losing the goal. That is why Meta sells them together.

A concrete scene helps. Ask it to sort out an insurance claim, make sense of a lease, or build a working app:

Week 1: "This is a lease, and clause 3 worries me"
Week 3: "Draft a dispute response based on that lease"
→ Without memory: "Which lease?" and start over
→ With memory: clause 3 is already on the table

Coding task: refactor that breaks three tests
→ Without focus: repeat the same fix after the first build failure
→ With focus: switch approaches, ask, and carry through to verification

Meta says it measures memory with MemoryBench-style evals and focus with FocusBench plus SWE-Marathon for sustained software work. On its own charts, Wonder Full scores 57.5 on SWE-Marathon against 46.3 for Claude Opus 5.5. Treat that as vendor-measured direction, not an independent verdict.

Read the scoreboard as wins and losses

Collecting the announcement and model card:

DeepSWE 1.1          Wonder Full 79.0% (strength)
SWE-Marathon         Wonder Full 57.5 vs Opus 5.5 46.3 (Meta's own measure)
Artificial Analysis
  Intelligence Index Wonder Full 56 (up from 48 on Spark 1.3)
  Terminal-Bench 4.0 33% -> 57% (Spark 1.3 -> Wonder Full, AA measure)
Terminal-Bench 4.0   Below Opus 5.5 on the model card (weakness)
Health / some coding Tasks where rivals lead also exist

So this is not "best at every coding task." It leans toward carrying long software projects through, while rivals can win on sharp terminal tasks or specific domains. That split matters more than any single number.

It is the same structure our SWE-bench vs Terminal-Bench vs ProgramBench guide describes:

DeepSWE / SWE-Marathon → stamina on long software work
Terminal-Bench         → multi-step execution in a terminal
→ different exams, so don't add the scores together

One more claim deserves a careful read: out-of-distribution scaling. Meta argues that at the same in-distribution score, Wonder Full handles unfamiliar tasks better than the Spark 1.3 trend predicts, crediting a stronger base model. Interesting direction, still Meta's interpretation.

Price and availability

Muse Wonder
  → Meta Model API + Muse Code
  → API price: $4 input / $20 output per million tokens
  → Muse app rollout starts with subscribers

Muse Wonder Full
  → Muse Code subscription only
  → Power / Maximum Usage plans
  → Maximum offers 80x the Everyday limit, five-hour and weekly resets

Wonder is for trying, Full is for the hardest workflows, including long-running agents and advanced computer use. At $4/$20, the right question is not token price but cost per finished task.

CodeBridge mini lab: run three models on one job

Turn the briefing's action into an experiment. Fix everything except the model:

1. Fix one job (e.g. fixing five failing tests)
2. Keep harness, tools, and time limits identical
3. Run Muse Wonder / Claude / GPT and record:

   [ ] completion rate (test-pass criterion)
   [ ] retries (approach changes after blockers)
   [ ] clarifications requested from a human
   [ ] token usage (input / output / cache)
   [ ] cost and time per finished task
4. Compute cost per success

Also track the shadow of persistence:

- long runs in the wrong direction (goal drift)
- whether it checked in before consequential actions

The model card recommends human confirmation before consequential actions. For agents that work longer, knowing when to stop and ask matters as much as raw capability.

Conclusion: buy the finish, not the first answer

Old contest: who writes the slickest first snippet
New contest: who keeps goals, constraints, and decisions to the end

One short next step: pick the most boring job in your repo this week and run all three models on it identically. Completion rate, retries, and cost per success on your work will decide next month's tooling better than any scoreboard.

Further reading

References

Go deeper with a course

Long jobs are finished by structure, not by the model alone. Practice placing memory and focus inside harness, loop, and graph designs.