Claude Sonnet 5.5 Confirmed for Coming Weeks; Specs Unannounced

Anthropic confirms Claude Sonnet 5.5 and Haiku 5.5, but not a date​

The one confirmed fact comes from the Opus 5.5 launch. Anthropic presented Opus 5.5 as the first model in a new Claude 5.5 family and said Sonnet 5.5 and Haiku 5.5 would follow soon. Anthropic product executive Mike Krieger repeated this on X: "We're also launching Claude Sonnet 5.5 and Haiku 5.5 in the coming weeks. These models will come with many of the same improvements to performance, efficiency, and safety." KDnuggets' roundup of the launch also reports that Sonnet 5.5 and Haiku 5.5 are both confirmed to launch soon after, per the same announcement.

The more specific timing comes from TheWinCentral, which reports claims that Sonnet 5.5 is already being tested, that some partners have temporary access, and that it could ship next week or in the first week of October, with September 30 or October 1 named as likely dates. No other outlet has reported those dates or the partner testing, and Anthropic's own announcement doesn't mention either. Treat the window as a rumor with a plausible basis. "Coming weeks," said on September 22, fits a late-September or October launch, but it doesn't pin one down.

The pace itself has a record behind it. According to KDnuggets, citing TechCrunch, Opus 5.5 arrived two months after Opus 5, which launched July 24, 2026. Anthropic is iterating fast, so a Sonnet refresh within weeks of an Opus release would be consistent with how it has been operating.

The Sonnet 5.5 "leaks" mostly restate Claude Sonnet 5's spec sheet​

The circulating claims give Sonnet 5.5 a 1-million-token context window, a 128,000-token maximum output, $2/$10 per million input/output tokens and $0.20 per million cache reads. Every one of those numbers already appears in Anthropic's published documentation for Claude Sonnet 5. TheWinCentral acknowledges the overlap, noting a September 24 report that flagged it.

Anthropic's documentation says Claude Sonnet 5 supports the 1M token context window by default (1M tokens is both the default and the maximum; there is no smaller context variant), 128k max output tokens, adaptive thinking. Anthropic's current pricing table lists Sonnet 5 at $2 per million input tokens and $10 per million output tokens, with cache hits at $0.20 per million. That is lower per-token pricing than Sonnet 4.6's $3/$15. TheWinCentral reports that Anthropic made the $2/$10 introductory rate permanent in August instead of moving Sonnet 5 to the previously announced $3/$15.

So a rumor that Sonnet 5.5 will keep 1M context and $2/$10 pricing is really a prediction that nothing changes on those fronts. That's a reasonable guess for a point release, but it tells you nothing new. If Anthropic does keep those terms, the upgrade would come down to quality and efficiency, not price or capacity. That fits how Anthropic described the family, though Anthropic hasn't committed to it.

For comparison, here is where things stand today:

Adaptive thinking and effort levels explain the "no xhigh" rumor​

The only claimed change with any substance concerns reasoning controls. The reports say Sonnet 5.5 will have adaptive reasoning on by default, with Low, Medium and High thinking options and no "xhigh" tier.

Part of that already describes Sonnet 5. With adaptive thinking, the model decides how much to reason based on the request. On Sonnet 5 it is the default. Anthropic's migration notes explain that on Claude Sonnet 4.6, requests without a thinking field run without thinking. On Claude Sonnet 5, the same requests run with adaptive thinking. To turn thinking off, pass thinking: {type: "disabled"}. Sonnet 5 also dropped manual extended thinking with fixed token budgets. Requests that set budget_tokens now return a 400 error, and Anthropic points developers to adaptive thinking with the effort parameter instead.

Anthropic's model overview lists Sonnet 5's default effort as "high" and Opus 5.5's as "medium." The Opus 5.5 benchmarks show that higher tiers exist: most results were run at "max" effort and Terminal-Bench 4.0 at "xhigh." If Sonnet 5.5 really tops out at High, it would sit clearly below Opus in reasoning spend. That reading is our inference from the rumored option list, not something Anthropic has said.

The reason it matters in practice is cost. On Sonnet 5, max_tokens is a hard cap on thinking plus response text combined, so reasoning tokens are billed as output. For agentic coding sessions that run hundreds of turns, the effort setting often decides the bill more than the per-token rate does.

Claude Opus 5.5's efficiency gains set expectations for Sonnet​

Opus 5.5 is driving the attention because its main selling point was efficiency, and Anthropic has said Sonnet 5.5 will share "many of the same improvements." Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads at default settings. Its per-token prices dropped to $4 input and $20 output, 20% below Opus 5, and cache reads dropped to $0.20 per million, 60% below Opus 5. TechCrunch independently reported the price cut: output tokens will be charged at $20 per million tokens for Opus 5.5, compared to $25 for the previous model.

Anthropic credits only part of that 40% to the price cut. It says Opus 5.5 "costs less per token than Opus 5 and uses fewer tokens per task," and that the two effects together produce the saving. Anthropic's early-access customers told the same story: Optiver reported Opus 5 quality in about half the turns, time and output tokens on its agentic coding tasks, and Rogo reported about 60% fewer output tokens on its finance benchmark. These are vendor-selected testimonials, not independent measurements.

For Microsoft developers, the notable name in that list is GitHub. Chief Product Officer Mario Rodriguez said that in GitHub's testing across GitHub Copilot CLI and VS Code, Opus 5.5 "used among the fewest tokens and steps we measured," and solved more terminal tasks than Opus 5 in less than half the steps in VS Code.

If Sonnet 5.5 inherits even part of that token discipline at Sonnet prices, the savings would add up for long-running agents. Anthropic has not published any Sonnet 5.5 efficiency figure, so any estimate of how much it will save is guesswork for now. Anthropic's commitment covers the categories of improvement: performance, efficiency and safety. It hasn't committed to a number for any of them.

Microsoft Foundry is where this reaches Azure shops​

Anthropic's current Claude models are sold inside Microsoft's own AI platform, and that's the main reason this matters to WindowsForum readers. Anthropic's Sonnet 5 documentation lists Microsoft Foundry among its launch platforms, alongside the Claude API, AWS and Google Cloud. KDnuggets reports that Opus 5.5 is live now on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Model IDs on Foundry match the Claude API aliases, for example claude-sonnet-5 and claude-opus-5-5.

Billing on Foundry works differently from the direct API. According to Anthropic's pricing documentation, Claude in Microsoft Foundry bills through the Azure Marketplace in Claude Consumption Units (CCUs). Anthropic rates token usage in USD at the standard per-model rates, applies any negotiated discount, converts the total to CCUs at a fixed $0.01 each, and reports usage to Azure hourly. Your Azure bill shows one CCU line item, and Azure Cost Management shows aggregated CCUs, not per-model detail. Deployments using the US Data Zone Standard type, which keeps inference inside the United States, carry a 1.1x multiplier on all token prices.

Whether Sonnet 5.5 will arrive on Foundry at the same time as the Claude API hasn't been announced. Anthropic did ship both Sonnet 5 and Opus 5.5 on Foundry at launch, so simultaneous availability is the likely pattern. It isn't a commitment.

Sonnet 5's tokenizer change is the budgeting lesson to carry forward​

One more detail from the Sonnet 5 launch matters for anyone planning a Sonnet 5.5 migration: per-token prices can mislead. Sonnet 5 introduced a new tokenizer that produces about 30% more tokens than Sonnet 4.6 for the same text. Anthropic says this means "the cost of an equivalent request does not drop in direct proportion" to the lower per-token price. The same 1M-token window also holds less text, and max_tokens limits tuned for Sonnet 4.6 can truncate output on Sonnet 5.

Nobody has said whether Sonnet 5.5 changes the tokenizer again. The earlier migration shows why to recount tokens, not assume them, when a new model ID appears. Enterprise consultancy Caylent made the same point about Sonnet 5, writing that it still needs task-level validation against real codebases before anyone treats it as a straight upgrade.

What this means for you​

Keep your current plans on Sonnet 5 and get ready to test Sonnet 5.5 when it ships. Don't budget or architect around rumored numbers. If you run Claude through the API or Microsoft Foundry, nothing forces a change today, and Sonnet 5 remains a supported, published target. Anthropic's model overview lists Sonnet 5's retirement as "not sooner than June 30, 2027."

• Base cost estimates and integrations on Sonnet 5's published terms: 1M context, 128K output, $2/$10 per million tokens and $0.20 cache reads.
• Treat September 30 or October 1 as an unconfirmed report. Anthropic's only commitment is "the coming weeks."
• When Sonnet 5.5 appears, recount representative prompts with the token counting endpoint and recheck max_tokens limits before switching production traffic, as the Sonnet 5 tokenizer change showed.
• Test the effort setting alongside the model. Opus 5.5 customers reported their biggest savings at lower effort levels, and Sonnet 5 defaults to "high."
• Foundry customers should confirm the new model ID in the Foundry catalog and watch CCU consumption in Azure Cost Management after switching, since the Azure bill doesn't break usage out by model.
• Remove budget_tokens and non-default temperature, top_p or top_k values now if you're still on older Sonnet integrations. Sonnet 5 already rejects them with a 400 error.

Anthropic has confirmed Sonnet 5.5 is coming, and it has tied the whole 5.5 family to the efficiency story that made Opus 5.5 notable: fewer tokens per task, lower cost per job. Everything past that, including the date, the price and the thinking tiers, is either Sonnet 5's current spec sheet or unconfirmed chatter. The first real decision point arrives with Anthropic's Sonnet 5.5 model page. When it goes live, compare its pricing, effort options and tokenizer notes against the Sonnet 5 baseline above, then run your own workloads before moving production traffic.