Anthropic released Claude Sonnet 5.5 on September 28, 2026, giving its faster Sonnet line a default 1-million-token context window, up to 128,000 tokens of output and a set of capabilities aimed directly at long-running, tool-using AI workflows. The model is available as claude-sonnet-5-5, and Anthropic positions it as the best balance of speed and intelligence in the Claude family. Anthropic's launch announcement says it runs more than 30% faster than Sonnet 5 while retaining the same $2-per-million input-token and $10-per-million output-token headline pricing.
The release is especially relevant to teams building research agents, browsing systems and automated workflows because the headline context expansion is only half of the story. Sonnet 5.5 also changes how several existing agent integrations behave. Anthropic's own documentation identifies five breaking changes for applications moving from Sonnet 5, including changes to thinking between tool calls, forced tool use, thinking-block handling and computer-use compatibility.
One million tokens becomes the default working context
Claude Sonnet 5.5 exposes a 1-million-token context window without requiring the older context-window beta mechanism, while maximum output reaches 128,000 tokens.
For agent systems, the practical significance is not simply that a single prompt can be unusually long. Long-running workflows accumulate state: source documents, search results, tool responses, intermediate analyses, user constraints, code, screenshots and the history of earlier actions. A larger working context gives the model more room to preserve that information as the task unfolds.
That can matter in research workflows where an agent must compare many sources before writing, in coding sessions spanning large repositories and in enterprise automation where the model has to preserve instructions while moving through several systems. The 1M window does not guarantee that every token will be used equally well, but it raises the ceiling on how much task state can remain available to the model at once.
Anthropic is positioning Sonnet 5.5 for fast professional and agentic work
Anthropic describes Sonnet 5.5 as a clear upgrade over Sonnet 5 and a faster, lower-cost complement to Opus 5.5. The company says it is strongest on well-scoped everyday tasks, bug fixing and the creation of polished documents, slides and spreadsheets.
In Anthropic's own evaluations, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0, compared with 10.3% for Sonnet 5, and comes within two points of Opus 5.5 on GDPval-AA. These are vendor-reported benchmark results and should not be interpreted as universal measures of real-world agent reliability, but they help explain where Anthropic is positioning the model: below Opus for the most demanding judgment-heavy work, but optimized for much faster iteration across everyday professional tasks.
Anthropic also says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and can cost up to 30% less per task because it often needs fewer tokens to accomplish the same work, even though list token prices are unchanged.
Long context changes the economics of research agents
A research agent typically operates as a loop rather than a single answer. It searches, opens sources, extracts evidence, compares claims, notices gaps, searches again and eventually synthesizes a result. Every additional step creates information that can become relevant later.
A larger context window can reduce the pressure to aggressively summarize or discard earlier material. An agent can potentially retain more original source text, tool results and decision history while continuing the investigation. That is useful for traceability because compression itself can remove qualifiers, dates or contradictory evidence that later turns out to matter.
But one million tokens should not be confused with one million tokens of guaranteed attention. Retrieval, source ranking, context management and explicit provenance remain important. Giving an agent a larger evidence pile does not automatically teach it which evidence deserves the most weight.
Browsing agents gain room for longer trajectories, not permission to trust everything they read
The same distinction matters for browser automation. Anthropic's documentation now separates browser use from full computer use: browser tools are intended for tasks that remain inside webpages, while computer use provides screenshot, mouse and keyboard control over a desktop environment.
For a browsing agent, additional context can preserve a longer sequence of pages, extracted facts and user requirements. That can make multi-site comparisons and extended research less dependent on repeated summarization. Yet browser content remains untrusted external input. More context can also mean retaining more irrelevant or adversarial material, so prompt-injection defenses and tool permission boundaries become more important rather than less.
The migration can break existing tool-using agents
Teams upgrading from Sonnet 5 should not treat the model ID as a drop-in replacement. Anthropic's Sonnet 5.5 overview explicitly lists five breaking changes. Among them, “forced tool use” configurations can now fail, thinking behavior changes between tool calls, thinking blocks become tied to the model and conversation, and the older computer-use interface is rejected on some platforms.
This matters because tool use is the control plane of an agent. A normal chatbot migration can often tolerate subtle differences in prose. An automated workflow cannot tolerate a tool call unexpectedly returning HTTP 400, a reasoning block becoming incompatible with a fallback model or an agent loop misreading a changed tool-use structure.
Production teams should therefore test complete trajectories rather than isolated prompts: the initial model response, tool selection, tool execution, tool results returned to Claude, subsequent reasoning and any fallback or routing behavior.
Computer use moves to the new toolset architecture
Anthropic's current computer-use documentation defines computer_toolset_20260801 as a client toolset containing 17 member tools, including screenshot, click, typing and zoom operations. A single toolset declaration exposes those capabilities, while the application executes the resulting actions inside an environment it controls.
For Claude 5.5 models on the Claude API and Google Cloud, the previous computer_20251124 tool is no longer accepted. Sending it produces a 400 invalid_request_error; integrations must migrate to computer_toolset_20260801 and update the agent loop for the toolset's member calls.
The platform distinction is important. Anthropic's documentation notes that older computer-use behavior can remain available on some other platforms and models, so developers should follow the compatibility matrix for the deployment environment rather than assuming one migration rule applies everywhere.
The new toolset changes how the agent loop is implemented
The migration is not only a renamed tool definition. Under the new architecture, Claude can return multiple member tool_use blocks in a turn. Each block identifies the member action by its name and carries toolset_name: "computer". The application must execute those actions and return results in the expected toolset structure.
That architecture is more expressive for agents capable of batching actions, but it means existing loops built around one older computer tool may need real application changes. Teams should therefore inspect both request construction and response handling before switching production traffic to Sonnet 5.5.
Tool search becomes increasingly valuable as agents gain more capabilities
Large agent systems can expose dozens or hundreds of tools. Loading every tool definition into the model's active context is expensive and can make tool selection harder. Anthropic's tool-search mechanism allows definitions to be deferred until Claude discovers that it needs them, while the API retains the full definitions server-side.
The computer and browser toolsets participate in that architecture through per-member deferred-loading configuration. This is an important companion to the 1M context story: a larger context window does not remove the incentive to manage context efficiently. The best agent architecture can combine a large maximum working memory with selective loading so that irrelevant tool schemas do not consume attention on every turn.
Adaptive thinking changes the balance between latency and reasoning
Anthropic's Sonnet 5.5 documentation emphasizes adaptive thinking and effort controls rather than treating reasoning as a simple on/off feature. The broader Claude 5.5 model family lets developers tune how much reasoning effort is spent on a task, creating a trade-off between speed, token consumption and deeper checking.
This is especially useful in agent workflows because not every step deserves the same amount of reasoning. Reading a straightforward tool response may require little deliberation, while deciding whether conflicting evidence is sufficient to trigger a consequential action may justify more effort.
The architectural opportunity is to allocate intelligence dynamically rather than run every step at maximum reasoning depth. Combined with faster generation, this can make Sonnet 5.5 attractive as the high-frequency model inside workflows where an Opus-class model would be unnecessarily expensive for every intermediate action.
Pricing remains $2 in and $10 out, with cheaper cache reads
Anthropic prices Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million tokens and cache writes at $2.50 per million. Opus 5.5, by comparison, is listed at $4 per million input tokens and $20 per million output tokens.
That makes context management economically important. A million-token context window is a capacity limit, not an invitation to resend a million fresh tokens on every turn. Prompt caching, selective retrieval and deferred tool loading remain essential when long-lived agents repeatedly operate over large bodies of information.
For AI search, the important shift is from answer generation to persistent evidence handling
Search-oriented AI systems are increasingly expected to do more than issue one query and summarize ten links. They may need to investigate follow-up questions, navigate websites, compare primary and secondary sources, retain prior evidence and use tools to verify details.
Sonnet 5.5's combination of a 1M context window, faster generation and tool-oriented architecture makes those longer trajectories more practical. The model can potentially preserve more evidence across the research session while tools provide access to information that was never present in training data.
The limiting factor then shifts from “can the model fit these documents?” toward “can the agent retrieve the right evidence, preserve provenance and know when to act?” That is a more demanding engineering problem than context length alone.
The migration checklist matters as much as the benchmark chart
Claude Sonnet 5.5 is easy to summarize as a faster Sonnet with one million tokens of context. For developers already operating agents, that summary misses the most actionable part of the release. Existing tool configurations and reasoning loops need to be checked before migration.
Teams should verify tool-choice behavior, thinking-block handling, cross-model fallback behavior and computer-use compatibility in the exact platform they deploy on. They should also measure full workflow cost rather than token price alone, because Anthropic's main efficiency claim is that Sonnet 5.5 completes work with fewer tokens and greater speed, not that its list price is lower than Sonnet 5.
The larger story is that context windows, browsing and tool use are converging into one agent infrastructure problem. A million-token model can remember more of the job, but useful autonomy still depends on disciplined retrieval, controlled tools, compatible agent loops and clear boundaries around what the model is allowed to do. Claude Sonnet 5.5 raises the ceiling on the first part while making the engineering requirements around the rest impossible to ignore.