Claude Sonnet 5.5 appears to be nearing release, based on a wave of developer posts on X and traces found inside Claude Code. The original report says the model may launch as soon as Tuesday afternoon U.S. time and describes it as much closer to Opus 5.5 in performance while carrying a far lower price.
According to the report, Sonnet 5.5 has already entered the final pre-release stage and is being rolled out through gray testing in Claude Code. Several early demos and public test results shared by developers claim the model beats the newly released GPT-6 Sol in coding and agent-related tasks, and in some cases comes close to OpenAI’s flagship GPT-6 Astra. A leaked pricing sheet cited in the article puts input at $2 per million tokens, output at $10 per million tokens, and cache reads at $0.20 per million tokens.
Model identifiers reportedly surfaced inside Claude Code
On Sept. 22, when Anthropic released Opus 5.5, the company also wrote in its documentation that Sonnet 5.5 and Haiku 5.5 would arrive in the coming weeks. The report says few expected Sonnet 5.5 to show up this quickly.
Developer @Mr_Salio wrote on X that Claude Sonnet 5.5 had been spotted in artifacts. The report says a dedicated model configuration identifier, "claude-sonnet-5-5," is already visible. In the same configuration file, the article says, there are also entries for "claude-opus-5-5," "claude-sonnet-5," and a separate identifier, "claude-fable-5.1."
The article interprets those front-end code entries as a sign that the model has already been deployed server-side and is waiting for a switch to be turned on. Developer @notjazii later added that Sonnet 5 had quietly been routed to version 5.5 for stealth testing about a week earlier.
The report also describes an informal way users have tried to check whether their accounts were included in the gray rollout. In Claude Code, with web access and memory turned off, users ask a question about "tibo" and the Codex reset meme. According to the article, older versions may respond inaccurately, while accounts routed to 5.5 give a more precise explanation. Several users then posted test results saying some accounts had already been switched over.
Early tests focus on UI generation, coding, and agent tasks
Most of the discussion around Sonnet 5.5 centers on code generation, front-end UI and animation output, and more advanced agent-style work.
Developer @notjazii shared a comparison described as "four mainstream models generating motion effects from the same prompt." The task asked each model to produce complex UI animation effects. In the wording quoted by the report, Sonnet 5.5 "really rubbed both OpenAI models (Sol and Astra) into the ground." The article says the visual results point to stronger coding intuition and better aesthetic judgment from Sonnet 5.5.
One reaction cited in the report said Claude has long been stronger in UI and animation, but if the output was produced entirely through code generation, the result was still hard to believe.
Developer @chetaslua focused on day-to-day coding ability and what the article calls coding feel. Using the same three prompts, he ran six parallel tests, with each model generating one output and no manual edits applied afterward. The figures cited in the report say Sonnet 5.5 produced three files ranging from 90 KB to 131 KB and hit the 90-minute limit in all runs. GPT-6 Sol, by comparison, generated files between 20 KB and 29 KB and took 9 to 12 minutes.
@chetaslua was quoted as saying: "Go use Claude Code now. Sonnet 5.5 in Ultracode mode is insanely fast. It’s extremely efficient, and it brings back that very natural human feel from the Sonnet 4 era." In his view, Sonnet 5.5 clearly outperformed GPT-6 Sol at the same price point and under the same prompts.
The report goes further. It cites @srikanthvaluri and @buildwithrajath as saying that after reviewing early tests, Sonnet 5.5 performed well above expectations. While GPT-6 Astra still held a slight edge in polish, they said Sonnet 5.5 had effectively pulled level with Astra overall. In some advanced coding and agent tasks, the article says, it also moved close to Astra.
Leaked pricing targets the low-cost, high-frequency segment
Pricing is the other major reason the model has drawn attention. The article lists the following numbers:
- Input: $2 per million tokens
- Output: $10 per million tokens
- Cache reads: $0.20 per million tokens
In the report’s framing, that would place Sonnet 5.5 directly in the low-cost, high-usage bracket. The article says OpenAI priced GPT-6 Sol and Luna at less than half the level of GPT-5.6 to capture high-frequency use cases. If Anthropic can offer performance close to Opus 5.5 at this price, the company would be competing in the same band with a stronger product.
The report quotes Rajath Gowda as saying: "Sonnet 5.5 is becoming a big problem for OpenAI. If it really is close to Opus 5.5, Anthropic is about to make frontier capability extremely cheap. That’s what OpenAI should be most worried about."
From there, the article argues that Anthropic’s product strategy is becoming clearer: Opus 5.5 holds the high end, while Sonnet 5.5 is positioned to take the mid-tier and everyday development market through aggressive pricing.
Timing speculation centers on OpenAI DevDay
The release window has become part of the story. The article notes that Anthropic launched Claude Opus 5.5 on the same day OpenAI released GPT-6 Sol and Luna, and says Opus 5.5 delivered stronger performance than its predecessor while cutting cost by 40% and improving speed by 30%.
The report also lists several recent Anthropic milestones over the past two weeks: Claude reportedly solved the nine-loop amplitude problem in physics and set a world record, discovered an unknown biological enzyme system, and launched Claude Marketplace with more than 2,000 plugins.
Against that backdrop, the article contrasts OpenAI’s recent internal model disputes with separate incidents it describes as AI leaking user images and a public researcher GitHub token cheating episode. OpenAI’s annual developer conference, DevDay, is scheduled for Tuesday, Sept. 29. The Sonnet 5.5 gray-testing chatter spread widely just ahead of that date.
Several developers cited in the report suggested the model could be released next week, possibly on DevDay itself. The article also mentions speculation that a major teaser could appear at 2 p.m. New York time tomorrow, though there has been no official confirmation.
The report links rapid iteration to Anthropic’s data flywheel
As for why Anthropic may be moving so quickly, the article points to reinforcement learning, post-training workflows, and what it describes as a high-quality data flywheel.
When asked how Anthropic was iterating at this pace, @yacineMTB speculated that the core factor was an extremely high-standard data flywheel. The report then cites newly exposed internal metrics and says Anthropic has pushed that system to a level that rivals would find hard to ignore:
- 26% of core R&D work is already led by AI, specifically earlier Claude models
- More than 80% of internally produced code is written autonomously by large models
The article says this goes beyond cost reduction and efficiency gains, arguing that recursive self-improvement is beginning to show up in practice. It also says Anthropic estimates Claude could become fully automated as early as 2027, which would accelerate iteration even more.
Anthropic has not publicly confirmed Sonnet 5.5’s launch date, final pricing, or full capability range. For now, the information in circulation comes from screenshots, demos, and test results posted by developers on X, and the claims remain subject to official verification.
References and byline
The original article lists the following references:
- https://x.com/notjazii/status/2103884749202993411
- https://x.com/notjazii/status/2103884167104831573
- https://x.com/chetaslua/status/2104153687077904613
- https://x.com/chetaslua/status/2104259190432903599
- https://x.com/Mr_Salio/status/2104210803545030792
- https://x.com/Mr_Salio/status/2104178854633849145/
The report says the piece originated from the WeChat public account "新智元," written by Aeneas 大卫.