Claude Sonnet 4.6 is Anthropic’s February 2026 model for coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It is still useful, but the important label is not “newest Sonnet”; it is Active (legacy). Existing workflows can keep using it, while new systems should compare it with Sonnet 5 before standardizing.
The decision anchor is a lifecycle/price mismatch. Anthropic’s Sonnet 4.6 model page, checked September 20, 2026, lists 1M context, 128K output, and $3 input / $15 output per million tokens. The same page recommends migrating to Sonnet 5, listed at $2/$10. It also says Sonnet 4.6 is not scheduled to retire before February 17, 2027 on Anthropic-operated platforms. “Still available” does not mean “best new default.”
Key takeaways
Sonnet 4.6 targets coding, computer use, long-context work, agent planning, documents, and design.
The API model ID is
claude-sonnet-4-6; the page lists 1M context, 128K output, adaptive thinking with extended deprecated, and text/image input.API pricing is $3/M input and $15/M output, with cache-write/read rates and a 50% Batch API discount.
Anthropic reports improved computer use and Claude Code preference over Sonnet 4.5 in early testing; those are vendor studies, not a public unified rerun.
Anthropic labels 4.6 Active (legacy), recommends Sonnet 5, and gives a retirement floor of 2027-02-17 for its operated platforms. Bedrock and Google Cloud differ.
Claude Sonnet 4.6 at a glance
The Sonnet 4.6 model resources, prompt collection, and review collection keep the model-specific source trail. They do not grant Anthropic API credits, Claude subscription access, or browser availability.
| Question | Current answer | Boundary |
|---|---|---|
| Model role | Sonnet’s February 2026 upgrade for code, tools, and knowledge work | Positioning is not a success guarantee. |
| Model ID | claude-sonnet-4-6 | Bedrock uses a different provider ID. |
| Context / output | 1M / 128K tokens | API ceiling; clients can compact earlier. |
| Input / output | Text and images / text | Audio and video are not listed. |
| Thinking | Adaptive; extended thinking deprecated; high default effort | UI/provider controls may differ. |
| API rates | $3 input, $15 output per million; cache read $0.30 | Cache writes, batch and subscriptions differ. |
| Status | Active (legacy); migration to Sonnet 5 recommended | Legacy means no ongoing updates. |
| Retirement | Not sooner than 2027-02-17 on Anthropic-operated platforms | Bedrock/Google Cloud set their own dates. |
What changed, and what the benchmark rows do not prove
Anthropic describes a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It reports that Claude Code users preferred Sonnet 4.6 over 4.5 about 70% of the time, and over Opus 4.5 about 59% of the time in early testing. It also reports better resistance to prompt injection than 4.5 and similar behavior to Opus 4.6. These are Anthropic’s evaluation and safety claims; the complete task mix and sampling are not a public independent rerun.
The Sonnet 5 overview is the current-family comparison. Opus 4.7 and Opus 4.8 cover higher-cost routes and their own lifecycle. Do not combine their benchmark rows with Sonnet 4.6’s OSWorld-Verified or customer reports into one ranking.
Access and pricing without mixing products
Claude API: use
claude-sonnet-4-6where the organization and provider still expose it. Record model ID, effort, cache mode, input/output tokens, retries, and accepted-task results.Claude app and Cowork: Anthropic announced 4.6 as the default for Free and Pro at launch, but current product pickers and plans can change. Check the live account.
Claude Code/Desktop: CLI, Desktop and app selectors can differ. A community report saw 4.6 disappear from one CLI view while an explicit
/model claude-sonnet-4-6route remained available.Cloud providers: Bedrock, Google Cloud and Microsoft Foundry have provider-specific IDs and lifecycle schedules.
Cost: $3/$15 is API pricing, not a Claude subscription conversion. Cache rates, batch discounts, tool use and retries alter the effective cost.
Choose a scenario and run a migration check
| Situation | Start with | Check before committing |
|---|---|---|
| Existing Sonnet 4.6 API integration | Re-run a fixed regression sample | Model ID, provider route, cache mode, tests and retirement notice |
| Broad PR review or browser smoke tests | Compare 4.6 with Sonnet 5 on identical tasks | Findings, pass rate, recovery, latency and cost per accepted result |
| Deep architectural tracing | Add Opus as a specialist | Multi-hop findings, review edits, tool calls and premium cost |
| Long PDF or contract workflow | Use 4.6 or Sonnet 5 on a reversible sample | Page coverage, table/footnote fidelity, citations and human validation |
| New production system | Benchmark Sonnet 5 first, keep 4.6 as fallback | Lifecycle, provider availability, output quality and rollback path |
The Reddit agentic PR report is a useful workload-shaped anchor: 10 sessions per model, High effort, a 29-file/~4K-line PR, and 10 browser QA sessions. Sonnet found more review issues on average (9 vs 6), while both passed a 7-step browser flow; Sonnet cost about $0.24/run versus Opus $1.32. It is a routing hypothesis for one team, not a general score.
For browser-first work, the AI browser guide and Tabbit Browser overview explain a separate runtime. The Tabbit pricing page is not Anthropic API pricing.
What users and reviewers actually establish
The PDF report describes a 90-page PDF conversion that the author preferred over GPT-5.5. The comments add the important limitation: clean Markdown can silently drop rows, merge cells, alter numbers, or lose footnotes. Use an LLM for difficult documents only with page, table, and random-sample validation.
After Sonnet 5 appeared, a Reddit comparison reported a Terraform task where 4.6 followed variables better than Sonnet 5. It is a subjective, non-blinded observation, but it explains why an active legacy model can remain valuable for a narrow workflow.
YouTube adds review themes rather than controlled scores. Developers Digest has chapters on computer use, benchmark caveats, adaptive thinking, GUI safety, and 1M beta limits. Binary Verse AI has chapters on price per success, the 1M beta trap, and deployment safety. Treat these as checklists, not proof of a universal winner.
Tabbit’s role and limits
If work begins with pages, tabs, files, screenshots, or browser automation, Tabbit Browser can provide a separate browser-first runtime. It does not extend Anthropic’s context, change legacy retirement dates, or prove Sonnet 4.6 appears in your picker. This draft did not run an authenticated Sonnet 4.6 task, so verify the live label and effective limits with a small reversible workflow.
Tabbit BrowserThe CTA downloads Tabbit Browser; it is not a Claude entitlement. Keep browser costs, Anthropic subscription limits and API rates separate.
Unknown risks
Legacy drift: Anthropic says legacy models no longer receive updates and recommends active models; behavior and support expectations can change before retirement.
Provider divergence: Anthropic-operated platforms, Bedrock and Google Cloud can have different IDs, availability and retirement dates.
Long-context fidelity: 1M is a ceiling, not proof that every page, table, citation or instruction survives attention and compaction.
Benchmark transfer: vendor tests, customer quotes, YouTube demos and Reddit pipelines use different prompts, tools, effort and dates.
Computer-use safety: prompt injection, destructive actions and false success claims still require permission boundaries, logs and human approval.
Verdict
Claude Sonnet 4.6 is still a credible compatibility model for coding, document work and browser QA. Its strongest practical case is not “it beats every newer model”; it is that a fixed, familiar workflow may still produce accepted work at predictable quality. But Anthropic labels it Active (legacy), recommends Sonnet 5 at a lower API price, and gives only a retirement floor. New systems should benchmark Sonnet 5 first; existing systems should pin 4.6, test the replacement, and keep a rollback path.
The model page tells you what Anthropic exposes. Your regression sample tells you whether migration is safe.
FAQ
What is Claude Sonnet 4.6?
Claude Sonnet 4.6 is Anthropic’s February 17, 2026 model for coding, computer use, long-context reasoning, agent planning, knowledge work, and design. Its API ID is claude-sonnet-4-6.
What are Sonnet 4.6’s context and output limits?
The checked model page lists a 1M-token context window and 128K maximum output. It accepts text and image input and returns text; clients and providers can expose less.
How much does Claude Sonnet 4.6 cost?
The API page lists $3 per million input tokens and $15 per million output tokens, plus cache-write and cache-read rates. Batch API input and output receive a 50% discount. Claude subscriptions and Tabbit pricing are separate.
Is Claude Sonnet 4.6 deprecated?
Anthropic labels it Active (legacy), says users should consider migrating to Sonnet 5, and lists retirement not sooner than February 17, 2027 for Anthropic-operated platforms. Bedrock and Google Cloud set their own schedules.
Is Sonnet 4.6 better than Sonnet 5 or Opus?
There is no universal winner. Anthropic recommends Sonnet 5 as the current Sonnet route, while community tests found 4.6 useful for breadth-first review and browser QA. Opus can still earn its higher cost on deeper multi-hop analysis. Test your workload.
Can I use Sonnet 4.6 in Tabbit Browser?
This draft did not run an authenticated Tabbit Sonnet 4.6 task and makes no claim about picker visibility, effective context, latency, or availability. Check your live picker and test a reversible workflow; Tabbit pricing is separate from Anthropic API pricing.