Claude Sonnet 5 is Anthropic’s attempt to make capable agentic work economical enough for routine production. Released on June 30, 2026, it combines a million-token context window, stronger tool use, adaptive reasoning, and launch pricing below the previous Sonnet generation.
Anthropic calls it the company’s most agentic Sonnet. The qualifier matters. Sonnet 5 is the strongest model in the Sonnet line, but Anthropic still positions more expensive Claude models for the hardest agentic coding and enterprise work. The product case is not “Sonnet replaces everything.” It is that more workflows can reach acceptable quality without paying flagship rates on every request.
The API model ID is claude-sonnet-5. It launched across all Claude subscription levels, became the default model for Free and Pro users, and is available in Claude Code and the Claude Platform. Access, rate limits, and product features still depend on plan and region.
The core specifications
Sonnet 5 provides a one-million-token context window by default, with no smaller context variant, and supports up to 128,000 output tokens in the synchronous Messages API. Its reliable knowledge cutoff and training-data cutoff are both January 2026.
The model accepts text and image input and supports vision, PDF input, the Files API, prompt caching, Batch processing, and Anthropic’s server- and client-side tool ecosystem. Priority Tier was not supported at launch, an operational detail that matters for teams with strict latency commitments.
One million tokens is capacity, not a recommendation to send every available document. Long inputs cost money, take time, and increase the chance that low-value context competes with the evidence that matters. The model also uses a new tokenizer, so the same text may occupy a different number of tokens than it did on Sonnet 4.6. Anthropic describes the increase as content-dependent rather than a fixed conversion.
Production systems should count tokens with the selected model, retrieve only relevant material, and reserve space for reasoning, tool calls, and the final response.
Adaptive thinking is the default
Sonnet 5 enables adaptive thinking by default. Instead of requiring a developer to assign a fixed internal reasoning-token budget, the model decides how much thinking a request deserves within the overall output limit.
Anthropic pairs this with five effort levels: low, medium, high, xhigh, and max. High is the default in the API and Claude Code. Effort is a behavioral signal, not a guaranteed token quota. It can influence reasoning, answer length, tool use, and even the detail of function arguments.
That makes effort a product-control surface. Low or medium may be appropriate for classification, extraction, and predictable transformations. High is a reasonable starting point for general agent work. Xhigh or max belongs on difficult cases where measured gains justify extra latency and output consumption.
Higher is not automatically safer. A high-effort model can spend more time pursuing a mistaken premise or call unnecessary tools when the success criteria are vague. Teams need task-specific evaluations and maximum-cost boundaries.
What “agentic” means here
Anthropic highlights improvements in planning, reasoning, tool calling, coding, and knowledge work compared with Sonnet 4.6. Sonnet 5 can operate with browsers, terminals, code environments, files, and remote tools when those capabilities are supplied and authorized.
An agentic model should be judged by completed workflows rather than isolated answers. Useful measurements include:
- whether it selects the right tool and arguments;
- how often it recovers from partial failure;
- whether it preserves IDs and state across long sequences;
- the number of unnecessary calls or retries;
- the quality of its stop and approval decisions;
- total elapsed time and cost per accepted result.
Tool access does not grant universal competence. A browser can expose prompt injection. A terminal can magnify a destructive command. A long context can carry contradictory or untrusted instructions. Agent design still needs sandboxing, least privilege, structured tool schemas, and independent validation.
Price and the promotion window
Anthropic launched Sonnet 5 with promotional API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. Standard pricing begins September 1 at $3 input and $15 output per million tokens.
The temporary rate makes evaluation cheaper and partially offsets the new tokenizer during migration, but teams should not use it as a permanent budget assumption. Forecast September pricing before committing a workload. Include cache writes and reads, Batch discounts where applicable, tool charges, reasoning output, and retries.
The relevant number is cost per accepted task. A model at $2/$10 can be expensive if it repeatedly calls a slow external tool. It can also be economical if better planning removes human correction and failed attempts. Compare end-to-end unit economics rather than multiplying an average prompt by the advertised rate.
Where Sonnet 5 fits against larger models
Anthropic says Sonnet 5 at high effort can match Opus 4.8 on some tasks, including specific browsing and computer-use evaluations. That is not a claim of universal parity. Model rankings vary by workload, prompt, tools, and evaluation method.
Sonnet is the natural default when speed, cost, and strong general capability all matter: coding assistance, research synthesis, document workflows, customer operations, and tool-based automation at meaningful volume. More expensive models remain candidates for ambiguous architecture, difficult debugging, high-impact analysis, and agentic work where a small quality improvement outweighs the price.
A useful routing policy starts Sonnet 5 on work it can validate, escalates to a larger model after a failed rubric or detected complexity, and records why the escalation occurred. Routing based only on prompt length wastes both money and capability.
Safety is improved, not solved
Anthropic reports a lower overall rate of concerning behavior than Sonnet 4.6 in its evaluations. The company also says Sonnet 5’s automatic behavioral-audit score remains behind Opus 4.8 and Mythos Preview, and its dangerous cyber capability is lower than those higher-end models.
These findings should not be compressed into “safer than every predecessor” or “safe enough for autonomous use.” Safety depends on the deployed system: tools, data, permissions, monitoring, and human oversight. A model that is less capable at a dangerous task can still misuse an over-privileged tool.
For production agents, use explicit allowlists, confirmation for external writes, file and URL validation, spend caps, and audit logs. Treat content from websites and uploaded documents as untrusted data even when it appears inside a long context.
Practical evaluation before adoption
Build an evaluation set from real work rather than generic benchmark questions. Include successful examples, known edge cases, missing inputs, conflicting instructions, tool timeouts, and policy-sensitive requests. Run the same cases at several effort levels.
Score correctness, completeness, tool behavior, latency, token usage, and human review time. Inspect the tail: a model that is fast on average but occasionally loops for ten minutes may not meet an operational objective.
If moving from Sonnet 4.6, use the new model ID in a canary, recalculate tokens, remove unsupported manual thinking parameters, and test refusals as a normal response state. Do not shift all traffic because the request schema looks familiar.
An approval-based media operation through MCP
Sonnet 5 can serve as the orchestration layer around a specialized media tool. Imagine a user asks Claude to add an approved WAV narration to a silent MP4. Claude can inspect filenames and metadata, explain the planned operation, request permission to transfer the files, call an authorized MCP tool, monitor the asynchronous job, and verify the returned media properties.
Medux is one external service that can perform such an operation; it is not native to Claude Sonnet 5. The Medux Claude MCP audio-track tutorial shows the operational sequence: obtain upload URLs, upload video and audio inputs, create the task, poll its status, and download the final MP4.
A safe Sonnet 5 workflow would keep control points explicit:
- confirm that the user owns or may process both files;
- inspect duration, format, and intended output before upload;
- present the selected inputs and estimated operation for approval;
- submit once with an idempotency key and store the task ID;
- poll at a bounded interval rather than asking the model to improvise retries;
- verify container, duration, and audio presence after download;
- require a person to review synchronization and publication rights.
The model’s stronger planning can make the sequence easier to coordinate. It does not replace the media engine, grant rights to the assets, or make the output correct by declaration. The best use of Sonnet 5 is to connect clear human intent to bounded, observable tools—and to stop when the evidence or authority is missing.