Leaked Claude Sonnet 5.5 at $2/M Tokens Reportedly Rivals GPT-6 in Coding Tests
新智元 · wechat · 2026-09-28
Just days after Opus 5.5 shipped, Sonnet 5.5 appears to be next. Since last night, word has spread on X that Sonnet 5.5 will launch soon, possibly Tuesday afternoon US time. It is said to be closer to Opus 5.5 in performance but far cheaper, now in final pre-release stage and in gray-scale testing inside Claude Code.
Leak evidence
Developer @MrSalio says an artifact already contains the model config identifier "claude-sonnet-5-5," alongside "claude-opus-5-5," "claude-sonnet-5" and a mysterious "claude-fable-5.1." Front-end code shipping means the model is deployed server-side, waiting on a switch. @notjazii says Sonnet 5 was quietly routed to 5.5 for shadow testing about a week ago.
Benchmarks
- In a pure-JS front-end UI/animation showdown, @notjazii says Sonnet 5.5 "wiped the floor" with OpenAI's Sol and Astra, showing strong code intuition and aesthetic taste.
- @chetaslua ran 6 parallel runs with the same 3 prompts: Sonnet 5.5 produced three files of 90-131KB, all hitting the 90-minute cap; GPT-6 Sol produced 20-29KB in 9-12 minutes. He calls Sonnet 5.5 in Ultracode mode blazingly fast and says it recovers the natural "human feel" of the Sonnet 4 era.
- Leakers @srikanthvaluri and @buildwithrajath, synthesizing early tests, say Sonnet 5.5 far exceeds expectations; while GPT-6 Astra retains a slight edge in polish, Sonnet 5.5 matches Astra and approaches it in some advanced coding and agent tasks.
Pricing
Reported at $2 per million input tokens, $10 per million output tokens, and just $0.20 per million cache reads. Developer Rajath Gowda says bluntly: "Sonnet 5.5 is becoming a big problem for OpenAI." Anthropic's strategy is clear: Opus 5.5 holds the high end while Sonnet 5.5 takes the mid-tier and daily dev market on price-performance.
Timing and context
Next Tuesday (Sept 29) is OpenAI's annual DevDay, and several big accounts hint the model lands that day. In the past two weeks Anthropic has claimed Claude computing a physics nine-loop amplitude to a world record, discovering an unknown enzyme system, and launching a Claude Marketplace with 2,000+ plugins. Insiders speculate Anthropic has nailed its RL/post-training pipeline, powered by a high-standard data flywheel: 26% of core R&D work is already AI-led, over 80% of self-produced code is written by models, and Anthropic estimates Claude could be fully automated as early as 2027.
Related event: Claude Sonnet 5.5 leak: gradual rollout underway, launch imminent(20 posts)→
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a context-bloat experiment — MandeepKaur9407 · 2026-11-16
- Berkeley paper: LLMs answer with stale values even when the update is in context — ombre-ai · 2026-11-16
- Effective HTML: 4.2k-star agent skills turn wireframes into self-contained interactive prototypes — tom_doerr · 2026-11-16
- skills-gate: one gate skill fronts 148 agent skills at just 389 tokens — Distinct-Pie2389 · 2026-11-16
- Open-weight Karpo-Decision-31B tops Decision Index at 73.9, one forward pass per answer — CodeByPoonam · 2026-11-16
- shadcn boosts @uisdk: one-shot UI rebuild that's instantly E2E-testable by AI — shadcn · 2026-11-16