The everyday Claude model keeps Sonnet 5’s token prices, while Anthropic claims faster output and fewer tokens spent on typical work.
Listen
AI narration
12:25
0:00 / 12:25
AI SummaryGenerated from this article
Anthropic released Claude Sonnet 5.5 on September 28, 2026, maintaining Sonnet 5's per-token pricing of $2 per million input tokens and $10 per million output tokens while claiming up to 30% lower task costs through faster output generation and reduced token consumption on routine work. The model is positioned for well-scoped jobs like coding fixes and document production, with availability across Claude, Claude Code, the API, and major cloud providers. However, independent testing has not yet verified these savings across production workloads, and actual cost reductions depend on individual deployments and whether quality remains acceptable for each use case.
A faster model does not necessarily cost less to use. Anthropic’s pitch for Claude Sonnet 5.5 is that this one can: it retains Sonnet 5’s per-token prices but, the company says, generates output more than 30% faster and completes typical tasks with fewer tokens. Anthropic estimates savings of up to 30% per task in its own testing.
Introduced on September 28, 2026, Sonnet 5.5 replaces Sonnet 5 as the middle-tier model in the Claude 5.5 family. It is available across Claude, Claude Code, Anthropic’s API and major cloud platforms. Teams using Sonnet for coding agents and routine office work can try it now, though independent testing has yet to establish how much of Anthropic’s claimed improvement carries over to production workloads.
The distinction is between a cheaper task and a cheaper token. Anthropic is promising the former; it is not changing the latter.
Sonnet 5.5 Is Built for Defined Work, Not Every Opus Task
Anthropic positions Sonnet 5.5 as the model for well-scoped jobs: fixing bugs, working through everyday coding requests, and producing documents, slides and spreadsheets. Opus 5.5, released six days earlier, remains its choice for complex, open-ended work that demands sustained judgment.
An agent asked to make a contained code change may benefit more from quick iterations and economical tool use than from the strongest available reasoning model. A sprawling architecture review has a different failure cost: a cheaper first attempt offers little value if a second model must redo the work.
Anthropic says Sonnet 5.5 sometimes approaches or exceeds Opus 5.5 in its evaluations. It also says Opus remains clearly stronger on difficult, less-defined assignments in its own testing and feedback from external testers. Sonnet’s appeal is that more work may fit within the less expensive tier without requiring an escalation.
The comparison depends partly on how hard the model is allowed to work. Anthropic’s cost-versus-performance charts show Sonnet 5.5 drawing closer to Opus at higher effort settings, but sometimes at a similar cost per task. Developers choosing between them need to compare configurations and completed jobs, not model names alone.
Faster Output Does Not Mean a Lower Token Rate
Anthropic lists Sonnet 5.5 at the same rates as Sonnet 5:
Input: $2 per million tokens
Output: $10 per million tokens
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
Those are token prices, not a fixed price for a bug fix or a finished slide deck. Anthropic says the new model typically uses fewer tokens to do comparable work and, in its tests, costs up to 30% less per task. A job that consumes exactly the same mix of input, output and cached tokens as before would cost the same under these rates.
For agents, a bill can reflect repeated reasoning, tool calls and revisions instead of one answer. Anthropic says early testers saw Sonnet 5.5 batch tool calls more often than Sonnet 5, reducing the steps needed for some coding work. That observation appears in Anthropic’s launch material; it is not an independently established saving across all agent setups.
The speed figure needs similar care. “More than 30% faster” describes output generation compared with Sonnet 5, according to Anthropic. It does not promise that every Claude Code session or API workflow will finish 30% sooner. Waiting for a tool, reading a large repository, executing tests and responding to an external service can all dominate the elapsed time. A faster-generating model can still be valuable in an interactive session, but the end-to-end gain depends on where that session spends its time.
For teams comparing models, the useful unit is the cost of a successful task. Fewer tokens on an unsuccessful attempt can be a false economy; lower token use with the same accepted result is a meaningful improvement.
Anthropic’s Benchmarks Show Gains, With Limits
The largest number in Anthropic’s release is from Terminal-Bench 4.0, an agentic command-line coding evaluation. Anthropic reports 70.6% for Sonnet 5.5 versus 10.3% for Sonnet 5. On CursorBench 4.0, which evaluates coding agents on tasks drawn from Cursor sessions, it reports 55.5% for Sonnet 5.5, 34.1% for Sonnet 5 and 57.8% for Opus 5.5.
Those figures warrant attention, particularly the reported gap over the previous Sonnet model. They should not be read as independently verified performance in a reader’s repository. The scores and comparisons come from Anthropic’s release and supporting Sonnet 5.5 system card; the results were not independently reproduced for this article. Agent benchmarks also depend on their harness, tools, effort settings and task selection. A score from one setup is not a general probability that a model will complete any coding request.
The separation is much narrower in Anthropic’s knowledge-work results. Its GDPval-AA v2.1 scores put Sonnet 5.5 at 1,844 and Opus 5.5 at 1,846, compared with 1,449 for Sonnet 5. The models can be close on some evaluated work, as Anthropic claims. That does not erase its stated preference for Opus on complex, open-ended assignments, and a two-point difference by itself does not establish practical equivalence.
“Near Opus” depends on the work being measured and how much effort each model receives. The case for Sonnet 5.5 is strongest when its quality is sufficient at a lower total cost. Even a large benchmark lead cannot settle that calculation for an individual deployment.
Claude, Claude Code, the API and Clouds Get Day-One Access
Anthropic says Sonnet 5.5 is available in Claude products, including Claude Code, and on the Claude Platform. It also lists access through Amazon Web Services, Google Cloud and Microsoft Azure. Developers do not have to wait for a separate announced cloud launch before considering an evaluation.
A team can compare the model in the environment it already uses, with its existing prompts, tools and acceptance checks. Access alone does not guarantee identical quotas, configurations or results across providers; those details matter when comparing runs.
For Claude users, the Sonnet tier is being upgraded for the defined coding and document work Anthropic says it handles best. API customers can use existing task logs to check whether the new model reduces tokens, time to a usable result or the number of retries without sacrificing quality.
New Cyber Fallbacks Could Affect Security Work
Sonnet 5.5 is the first Sonnet release Anthropic says comes with both cybersecurity safeguards and fallback behavior similar to those used for its more capable models. Sonnet already had cyber protections: Anthropic’s Sonnet 5 announcement said they were enabled by default.
The new consideration is what happens when the protections intervene. Anthropic says Sonnet 5.5’s cybersecurity capabilities are comparable to those of Opus 5, prompting the stronger safeguard-and-fallback approach. Some affected requests may be handled through a fallback instead of by Sonnet 5.5 itself. Anthropic says the measures target a narrow set of high-risk requests and that routine software development should be unaffected. Its biology safeguards remain the same as Sonnet 5’s.
A workflow involving vulnerability research or other sensitive cyber tasks may behave differently from a standard coding workflow, even if both use the same model entry point. A fallback can change which model completes a request, complicating comparisons of quality, latency and cost. The practical effect will depend on the request and applicable access arrangements; the launch figures do not establish a universal fallback rate.
Developers evaluating Sonnet 5.5 for security work should record safeguard interventions alongside ordinary completion metrics. Otherwise, an internal test may appear to measure one model consistently when some requests followed a different path.
The Production Test Is Cost per Accepted Result
Sonnet 5.5 arrives with an economic proposition: the posted Sonnet rate is unchanged, while Anthropic claims the model can finish many jobs faster and with less token use. Its benchmark results suggest a substantial upgrade, but they are vendor-reported results, not a substitute for testing against an organization’s own tasks.
A useful trial would hold the job and acceptance criteria steady while measuring total input, output and cache usage, tool calls, elapsed time, retries and the quality of the finished work. That comparison may favor Sonnet 5.5 for contained tasks and Opus 5.5 for harder ones. It may also show which workloads, if any, approach Anthropic’s claimed 30% task-cost saving.
Frequently Asked Questions
3 questions
1
Does Claude Sonnet 5.5 Cost Less per Token Than Sonnet 5?
No. Anthropic lists Sonnet 5.5 at the same prices as Sonnet 5: $2 per million input tokens, $10 per million output tokens and $0.20 per million tokens for cache reads. Its claim of up to 30% lower cost per task comes from using fewer tokens in its testing. Savings will vary with the workload and the number of successful attempts needed.
2
Is Claude Sonnet 5.5 Faster for Every Coding Task?
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5. That measures output generation, not the full duration of every coding task. Claude Code and other agents may also spend time running tests, calling tools and reading files. Anthropic has not independently established a 30% end-to-end speed gain for every workflow.
3
Can Claude Sonnet 5.5 Fall Back to Another Model for Security Work?
Yes. Anthropic says Sonnet 5.5 has cybersecurity safeguards and fallback behavior that can affect a narrow set of high-risk requests. It says routine software development should be unaffected, but sensitive security workflows may need closer evaluation. Teams should check for safeguard interventions when comparing performance, since a request handled through a fallback may not reflect Sonnet 5.5 alone.
