Anthropic launches Claude Sonnet 5.5 with 30% faster output, lower task costs and stronger agentic coding
Anthropic has launched Claude Sonnet 5.5, the second model in its new Claude 5.5 family, positioning it as a faster and more economical alternative to Claude Opus 5.5 for everyday production workloads, coding, agents and professional work.
The headline improvement is not a lower API price per token. Sonnet 5.5 keeps the same $2 per million input tokens and $10 per million output tokens pricing as Claude Sonnet 5.
Instead, Anthropic says the model generates output more than 30% faster and typically requires fewer tokens to complete the same task, reducing the effective cost of many workloads by up to 30% per completed task.
The performance increase is also substantial in several agentic evaluations. Sonnet 5.5 reaches 70.6% on Terminal-Bench 4.0, compared with 10.3% for Sonnet 5, while approaching Claude Opus 5.5 on several professional-work and coding benchmarks.
The result is a model designed to occupy an increasingly important part of Anthropic's lineup: workloads where users need near-frontier performance, fast iteration and large-scale deployment without paying Opus-level inference costs.
··········
CLAUDE SONNET 5.5 AT A GLANCE
........
Metric | Claude Sonnet 5.5 |
Release date | September 28, 2026 |
API input price | $2 / 1M tokens |
API output price | $10 / 1M tokens |
Cache reads | $0.20 / 1M tokens |
Cache writes | $2.50 / 1M tokens |
Output speed vs Sonnet 5 | 30%+ faster |
Cost per completed task vs Sonnet 5 | Up to 30% lower |
Terminal-Bench 4.0 | 70.6% |
FrontierCode 1.1 | 46.2% at Max effort |
CursorBench 4.0 | 55.5% |
GDPval-AA v2.1 | 1,844 |
API model | claude-sonnet-5-5 |
........
The pricing distinction is important.
Anthropic has not cut Sonnet's nominal input or output token rates. The lower cost claim comes from efficiency: if Sonnet 5.5 can reach the desired answer using fewer generated tokens, fewer reasoning steps or fewer tool interactions, the total bill for completing the task falls even when the unit price remains unchanged.
This makes cost per completed task more useful than token price alone when comparing agentic models.
··········
SONNET 5.5 IS NOT JUST A FASTER SONNET 5
The largest changes appear in tasks where the model must do more than answer a single prompt.
Agentic coding, repeated tool use, document production and longer workflows depend on the model's ability to decide what to do next, avoid unnecessary actions and maintain a useful trajectory toward the requested result.
Sonnet 5.5 shows a particularly large improvement on Terminal-Bench 4.0, an evaluation focused on completing coding tasks through a terminal environment.
The model scores 70.6%, versus 10.3% for Sonnet 5.
That gap is much larger than the type of incremental benchmark movement normally associated with a same-tier model refresh.
The result also exceeds the 66.4% reported for Opus 5.5 under Anthropic's published evaluation configuration, although a single benchmark should not be interpreted as evidence that Sonnet 5.5 is generally more capable than Opus.
Anthropic continues to position Opus 5.5 as the stronger model for complex, ambiguous and open-ended work requiring sustained judgment.
Sonnet 5.5 instead appears optimized around a different target: high-quality execution when the objective is reasonably well specified.
That distinction matters for production agents.
An agent fixing a known software bug, transforming a structured dataset, updating documents, generating a presentation or executing a defined operational procedure does not necessarily need the maximum possible reasoning depth on every step.
It needs enough intelligence to complete the workflow reliably while minimizing latency, tokens and unnecessary tool calls.
··········
THE BENCHMARK GAP WITH OPUS 5.5 HAS BECOME SMALL IN SOME WORKLOADS
Anthropic's published numbers show that the distance between Sonnet and Opus is no longer consistent across tasks.
........
Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% |
FrontierCode 1.1 | 46.2% | 42.4% | 54.4% |
CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
GDPval-AA v2.1 | 1,844 | 1,449 | 1,846 |
AA-Briefcase v1.1 | 1,811 | 1,359 | 1,822 |
........
The professional-work results are particularly notable.
On GDPval-AA v2.1, Sonnet 5.5 scores 1,844 compared with 1,846 for Opus 5.5.
On AA-Briefcase v1.1, the scores are 1,811 and 1,822 respectively.
Those differences are small compared with the gap between Sonnet 5.5 and Sonnet 5.
The coding picture is more mixed.
Opus 5.5 retains a clearer advantage on FrontierCode 1.1, while Sonnet 5.5 comes much closer on CursorBench and leads in the reported Terminal-Bench result.
This is a useful illustration of why model rankings based on a single benchmark can be misleading.
Different evaluations reward different behaviors, and agentic systems are particularly sensitive to scaffolding, effort settings, tool configuration and the amount of time the model is allowed to spend solving the problem.
Anthropic itself notes that some Sonnet 5.5 results change depending on the selected effort level.
In FrontierCode, for example, increasing effort can sometimes introduce extra review activity or out-of-scope edits that the benchmark penalizes rather than rewards.
More reasoning is therefore not automatically equivalent to a better task-level result.
··········
THE REAL COST ADVANTAGE COMES FROM COMPLETING WORK WITH FEWER TOKENS
Sonnet 5.5 preserves the same API pricing as Sonnet 5:
$2 per million input tokens
$10 per million output tokens
$0.20 per million cache-read tokens
Anthropic nevertheless estimates that the new model can cost up to 30% less per task.
Consider a simplified workload where Sonnet 5 requires 100,000 output tokens across an agent's complete sequence of reasoning, tool calls, retries and final responses.
At $10 per million output tokens, those outputs cost:
100,000 / 1,000,000 × $10 = $1.00
If Sonnet 5.5 accomplishes the same task using 30% fewer output tokens, output consumption falls to 70,000 tokens:
70,000 / 1,000,000 × $10 = $0.70
The token rate has not changed, but the task costs 30 cents less.
Real workloads are more complicated because they combine input tokens, output tokens, prompt caching and potentially large amounts of reused context.
The principle remains the same: model efficiency can reduce inference expenditure without a headline price cut.
This is especially important for agents because an inefficient agent can multiply a small per-token difference across dozens of model calls.
··········
SONNET 5.5 COSTS HALF AS MUCH AS OPUS 5.5 FOR STANDARD INPUT AND OUTPUT TOKENS
The economic distinction inside the Claude 5.5 family is straightforward.
........
API pricing per 1M tokens | Sonnet 5.5 | Opus 5.5 |
Input | $2 | $4 |
Output | $10 | $20 |
Cache reads | $0.20 | $0.20 |
Cache writes | $2.50 | $5 |
........
For normal input and output tokens, Opus 5.5 costs exactly twice as much as Sonnet 5.5.
That creates a significant routing opportunity for companies deploying Claude at scale.
A system can reserve Opus for tasks requiring deeper judgment, difficult ambiguity or sustained reasoning while directing routine coding, document generation, extraction, transformation and structured agent tasks toward Sonnet.
If Sonnet 5.5 also consumes fewer tokens than Sonnet 5, its effective cost advantage becomes larger than the nominal price comparison alone suggests.
For workloads involving millions or billions of tokens per month, this can materially change the economics of using a stronger model as the default rather than as an exception.
··········
SPEED MAY MATTER AS MUCH AS MODEL QUALITY FOR AGENTIC SYSTEMS
Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5, making it the fastest Sonnet model the company has released.
Latency compounds in multi-step workflows.
If an agent makes one model call, a modest reduction in generation time may be barely noticeable.
If it performs 20 sequential reasoning and tool-use steps, delays accumulate across the entire execution chain.
A 30% improvement at the model layer therefore has the potential to translate into a substantial reduction in end-to-end completion time, particularly where the next tool call cannot begin until the previous model response is complete.
This is also why agent performance cannot be evaluated purely through intelligence benchmarks.
A production model sits inside a larger system where quality, latency, cost, tool-call frequency and failure rate interact.
A model that reaches almost the same final quality while completing the process faster and with fewer tokens can have a better production profile than a more capable model whose additional reasoning is unnecessary for the task.
··········
EFFORT LEVELS GIVE DEVELOPERS ANOTHER COST AND PERFORMANCE CONTROL
Sonnet 5.5 supports adjustable effort levels, allowing users to trade additional reasoning against latency and token consumption.
Anthropic sets Medium effort as the default in Claude Code and its applications, while the Claude Platform defaults to High.
Lower effort settings are intended for faster, more routine work.
Higher settings allow Claude to reason longer and perform additional checking before completing the response.
This creates another optimization layer beyond choosing between Sonnet and Opus.
A production system can potentially route workloads across both model tier and effort level.
A routine transformation may use Sonnet 5.5 at a lower effort level.
A more difficult coding task might remain on Sonnet but increase effort.
A genuinely ambiguous or high-value problem could then escalate to Opus 5.5.
The model family is consequently moving toward a more granular compute allocation model rather than requiring developers to make a binary choice between a cheap model and an expensive one.
··········
DOCUMENTS, SLIDES AND SPREADSHEETS ARE PART OF THE TARGET WORKLOAD
Anthropic is explicitly positioning Sonnet 5.5 beyond software development.
The company highlights improvements in producing documents, presentations and spreadsheets, alongside clearer writing and better visual understanding.
This matters because business-document generation increasingly resembles an agentic task rather than conventional text completion.
Creating a finished presentation can involve reading source files, identifying relevant information, building a structure, producing text, interpreting charts and images, applying formatting rules and revising the result after inspecting the output.
Spreadsheet work can similarly require multiple rounds of data inspection, calculation, formula generation and validation.
These workloads reward the same characteristics that matter in coding agents: persistence, tool use, state tracking and the ability to detect when additional work is necessary.
Sonnet 5.5's role is therefore broader than that of a lower-cost coding model.
Anthropic is effectively positioning it as a general production model for repeatable professional workflows.
··········
SAFETY REQUIREMENTS HAVE ALSO INCREASED WITH THE MODEL'S CAPABILITIES
The improvement in coding capability has consequences beyond conventional software development.
Anthropic says Sonnet 5.5's cybersecurity capabilities are sufficiently strong that the model is being released with cyber safeguards and fallback mechanisms similar to those used for the company's most capable models.
It is the first Sonnet model to receive that level of cyber-specific deployment protection.
Anthropic says the biological safeguards remain equivalent to those used for Sonnet 5.
The company also subjected Sonnet 5.5 to its automated behavioral audit covering roughly 1,850 scenarios, including tests related to misleading behavior, acting against user interests and cooperation with high-stakes misuse.
According to Anthropic's evaluation, Sonnet 5.5 matches or improves on Sonnet 5 on most of those measures.
The additional controls illustrate a broader trend in frontier-model development: improvements designed for legitimate agentic work increasingly overlap with capabilities that require stronger deployment safeguards.
··········
AVAILABILITY IS IMMEDIATE ACROSS ANTHROPIC AND THE MAJOR CLOUD PLATFORMS
Claude Sonnet 5.5 is available through Anthropic's products and the Claude Platform under the API identifier:
claude-sonnet-5-5
Anthropic is also making the model available through Amazon Web Services, Google Cloud and Microsoft Azure, allowing existing enterprise deployments to adopt it without moving their entire AI infrastructure to Anthropic's own platform.
Zero-data-retention availability is supported, continuing the deployment option already offered with other recent Claude models.
For developers migrating existing Sonnet workloads, Anthropic has also introduced configuration changes around thinking behavior, including a new between_tools setting for workloads that previously operated with thinking disabled.
Migration therefore should not be treated purely as changing the model name in production systems.
Teams using custom agent scaffolding, effort settings or thinking configurations should validate model behavior and token consumption under their existing workflows before switching traffic completely.
··········
SONNET 5.5 CHANGES THE ECONOMICS OF THE CLAUDE 5.5 FAMILY
The most important characteristic of Claude Sonnet 5.5 is not that it replaces Opus 5.5.
It does not.
Anthropic still describes Opus as the stronger option for complex, open-ended tasks requiring sustained judgment.
What Sonnet 5.5 changes is the point at which paying for Opus becomes necessary.
On several professional-work evaluations, the new Sonnet operates very close to Opus 5.5 while charging half the standard input and output token price.
Compared with Sonnet 5, it retains the same nominal API pricing while delivering substantially stronger agentic results, generating output more than 30% faster and potentially reducing total task costs through lower token consumption.
For developers, the practical decision therefore becomes increasingly workload-specific.
Opus 5.5 provides the higher ceiling. Sonnet 5.5 is designed to make much of that capability economical enough to become the default.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
datastudios.org
