You take freelance gigs, ship a dozen projects a month. Which coding model?
Used to be a two-horse race between OpenAI and Anthropic. Now Meta crashed in with Muse Spark 1.1, and coding is a three-way fight for the first time.
Each contender's hill
Muse Spark 1.1 (Meta, July 9): #1 on MCP Atlas, JobBench, Humanity's Last Exam, and Finance Agent V2. On the Harvey legal-agent benchmark it scores 20% win rate vs Claude Fable 5's 11%. 1M context, can drive a computer. Priced $1.25 / $4.25 per million tokens. Zuckerberg came out of a three-year X hiatus to announce it; partners are Replit, Cline, Box.
GPT-5.6 Sol (OpenAI, June 26): strongest breadth. Coding Agent Index 80 vs Fable 5's 77.2, Terminal-Bench 88.8 vs 83.1. Output $30 per million tokens.
Claude Fable 5 (Anthropic, June 9, restored globally July 1): deepest repo-level refactors, SWE-Bench Pro 80% vs Sol's 64.6%. Long-form and complex logic are its strength. Output $50 per million tokens, the priciest.
Spread the prices out: Muse Spark output $4.25, Sol $30, Fable 5 $50. For the same ten thousand lines, Meta's price is nearly an order of magnitude cheaper. Not "a bit cheaper" — it craters the floor.
The catches, stated plainly
Muse Spark is closed-source, no open weights — Meta moved from open-source evangelist to a dual track; the community can't modify it. And its flashy benchmarks are mostly from Meta and partners (Replit etc.); independent verification is still pending.
Fable 5 has baggage: shut down by a U.S. export-control order on June 12, restored globally only July 1. Enterprises should price that supply-chain risk in.
Sol has the most mature ecosystem but the pricey tier; you may not use all its breadth.
Verdict: no hedging
- Heavy agentic, multi-tool, tight budget → Muse Spark 1.1, strongest agent benchmarks and cheapest
- Deep repo refactors, complex long logic → Claude Fable 5, still the king there
- Want breadth, ecosystem, one vendor → GPT-5.6 Sol, the balanced pick
- Freelance boss → try Muse Spark first, insane value; but closed-source means accepting lock-in
Why this matters to you
A three-way fight is good news for coders: Muse Spark's entry drove prices down. Agentic coding used to cost $30–50 per million output tokens; now $4.25 gets you first-tier level.
But don't chase cheap blindly — wiring a few APIs vs refactoring a 100k-line legacy repo are different jobs. Know what your work looks like before picking, not what the leaderboard says.