Claude Sonnet 5.5 and Haiku 5.5 Are Coming in Weeks

Anthropic launched Claude Opus 5.5 on September 22, 2026, and used the announcement to name the next two models in the family. Claude Sonnet 5.5 and Claude Haiku 5.5 will follow “in the coming weeks,” the company said, bringing many of the same improvements to performance, efficiency, and safety.

Neither model has a release date, price, or benchmark score yet. For people choosing a model for everyday work, those details will determine the value of the upgrade.

Opus shows the direction Anthropic is taking: stronger results at lower operating cost. How much of that progress will reach Sonnet, its middle tier, and Haiku, its fastest and cheapest tier, remains to be seen.

Anthropic Has Confirmed the Models, Not Their Specifications

Sonnet occupies the middle and Haiku the fast, low-cost end of the Claude lineup; Opus is the premium model. Each tier serves a different buying decision, even within the same model family.

Anthropic says Sonnet 5.5 and Haiku 5.5 will carry many of Opus 5.5’s improvements in performance, efficiency, and safety. It does not say every improvement will transfer, that both models will ship on the same day, or that their gains will be equal.

There are no published results for either model to compare with its predecessor. Anthropic has not specified their token prices, context windows, generation speeds, or rollout plans in the Opus announcement. “Coming weeks” was the company’s stated plan as of September 22, not a dated availability notice.

Opus 5.5 offers evidence of what Anthropic has improved at the top of its range. The value of the two forthcoming models is still an open question.

Opus 5.5 Shows Why Efficiency Matters More Than a Score

The Opus 5.5 pricing details give the clearest indication of Anthropic’s priorities. Standard input costs $4 per million tokens, down from $5 for Opus 5, while output costs $20 per million, down from $25. Cache reads cost $0.20 per million tokens instead of $0.50. Anthropic says Opus 5.5 also uses fewer tokens per task, producing a 40% cost reduction on what it describes as typical workloads.

Those figures affect bills in different ways. A 20% reduction in input and output rates does not mean every job costs 20% less. Cache-read savings help workloads that repeatedly reuse stored prompt material; savings from using fewer tokens depend on the task. Anthropic also says Opus 5.5 generates output more than 30% faster than Opus 5 at default settings. None of these figures establishes what Sonnet 5.5 or Haiku 5.5 will cost or how fast they will run.

Anthropic reports a 66.4% result for Opus 5.5 on Terminal-Bench 4.0, an agentic coding evaluation, at its specified xhigh effort setting. That is a result for one model under one test setup. It tells us neither how the unreleased models will score nor whether either will suit shorter, routine requests.

The Opus launch also changes the comparison the next models will face. A cheaper premium tier gives users another option for difficult tasks. Sonnet and Haiku will have to justify their places through quality, latency, and cost on the work people assign them, rather than through lower token rates alone.

Sonnet 5.5 Will Be Judged by Cost per Completed Job

Sonnet’s middle-tier role makes its upgrade particularly consequential for developers who need capable responses across many requests but cannot justify using a premium model for every one. Price per million tokens tells only part of that story.

Consider a coding assistant that attempts a change, checks its work, and tries again when something fails. A less expensive model can produce a higher bill if it needs more attempts or longer outputs to finish reliably. Conversely, a Sonnet model that succeeds more often with fewer tokens could save money even if its posted rates barely change. This illustrates a purchasing decision; it is not a claim about Sonnet 5.5’s unreported performance.

Outside coding, a team processing documents or drafting customer-facing material needs to know how often the model produces usable work, how much checking it requires, and how quickly it responds. A benchmark improvement may help answer those questions, but it will not settle all three.

Anthropic has highlighted improvements to Opus 5.5’s communication, saying it puts important information earlier and writes more clearly than Opus 5. The company has not explicitly promised the same writing changes for Sonnet 5.5 or Haiku 5.5. If similar changes appear in Sonnet, they could reduce the effort needed to review its answers. That is something to test when the model ships, not a feature to count on now.

Haiku 5.5 Has a Different Standard to Meet

Haiku’s appeal is speed and low cost, particularly for applications that handle many requests or whose users notice each delay. A fast model might classify an incoming message or decide whether a harder request should go to another model. These are potential uses of Haiku’s tier, not announced Haiku 5.5 capabilities.

Speed can mean different things in those workloads. Time to the first response may matter most in an interactive interface; total processing time and throughput may matter more for a batch job. Opus 5.5’s reported improvement in output-generation speed tells us nothing definite about those measurements for Haiku 5.5.

The cache-read price is another figure worth watching. Anthropic cut Opus’s rate sharply, and repeated prompt material can make caching important in high-volume applications. But the company has announced no Haiku 5.5 cache-read price. Developers should not budget for an equivalent cut simply because the premium tier received one.

Haiku also needs to be accurate enough for the work assigned to it. If a fast first pass frequently has to be repeated by Sonnet or Opus, its apparent savings may disappear. Depending on the application, a small improvement in reliability could be worth more than a dramatic-looking speed gain.

Safety and Communication Claims Need Their Own Evidence

Anthropic says the two coming models will share many of Opus 5.5’s safety improvements. It has not published an evaluation of either model. The company reports stronger results for Opus 5.5 on its automated behavioral audit and says external evaluators tested Opus before release. Those findings cannot establish how Sonnet 5.5 or Haiku 5.5 will behave.

The same qualification applies to performance. “Many of the same improvements” does not promise identical benchmark gains, safeguards, or behavior under long and complicated instructions. Different tiers make different tradeoffs. Anthropic will need to show what each model gained and what it still cannot do reliably.

When the releases arrive, three checks will be more informative than the family resemblance:

• Compare full task costs. Include input, output, cache reads and writes, and any retries, not just the advertised input rate.
• Test the intended workflow. Measure response time and usable results on representative requests, alongside whatever benchmarks Anthropic publishes.
• Read each model’s safety evidence separately. An Opus evaluation cannot establish how a different model behaves in the same scenario.

Those checks may also show that switching is not worthwhile. A current setup that meets its quality and latency requirements needs a measurable benefit from a newer model.

Final Thoughts

Opus 5.5 makes a credible case that Anthropic is working on efficiency as well as capability: its published token rates are lower, and the company reports reduced token use on typical tasks. Sonnet 5.5 and Haiku 5.5 could bring that progress to a much broader set of decisions, but neither has demonstrated it yet.

Their prices, speed, and reliability on real jobs will be more useful than an Opus benchmark. Until Anthropic publishes those details, the September 22 announcement is a reason to watch the next two launches, not to rewrite a deployment budget.