Anthropic Claude Sonnet 4.6
by Anthropic
Near-Opus coding and computer-use scores at $3/$15 per million tokens
Claude Sonnet 4.6 is Anthropic's February 2026 mid-tier model, still active on the API at $3 per million input and $15 per million output tokens. It landed within roughly a point of the far more expensive Opus 4.6 on SWE-bench Verified and OSWorld-Verified, pairs a 1M-token context with 128K output, and cut false refusals on benign prompts by about 47x — the release that made Sonnet-class economics viable for agentic work.
Claude Sonnet 4.6 is the mid-tier Claude model Anthropic released on 17 February 2026, still listed as Active on the API as `claude-sonnet-4-6` with a tentative retirement no sooner than 17 February 2027. It pairs a 1M-token context window with up to 128,000 output tokens — 300,000 through the Batch API extended-output beta — at $3 per million input tokens and $15 per million output, the same price as Sonnet 4.5, and supports adaptive thinking, tool use, structured outputs and computer use. Its headline claim was frontier-adjacent quality at Sonnet economics, and the numbers back it: 79.6% on SWE-bench Verified against Opus 4.6's 80.8%, 72.5% on OSWorld-Verified against Opus 4.6's 72.7%, and 59.1% on Terminal-Bench 2.0. Anthropic reported that Claude Code users preferred it to Sonnet 4.5 about 70% of the time and to the more expensive Opus 4.5 59% of the time. The safety numbers moved further than the capability ones: harmlessness on violative requests rose to 99.38% from 97.89%, and over-refusal on difficult benign prompts fell from 8.50% to 0.18% — roughly a 47x reduction in false refusals — under ASL-3 deployment safeguards, which is what made it usable for security research and other legitimate-but-sensitive work that earlier models stonewalled. Independent aggregation ranks it #35 of 214 models with 89.9% GPQA and 79.2% MMLU-Pro, and it is weakest on reasoning-heavy evaluations, scoring 58.3% on ARC-AGI-2 against Opus 4.6's 68.8%. It runs on claude.ai, Claude Cowork, Claude Code, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. Claude Sonnet 5 has since superseded it, at introductory pricing of $2/$10 through 31 August 2026.
Product and platform teams standardising a coding assistant or computer-use agent on a mid-priced model, who need a pinned snapshot with predictable behaviour and cannot justify Opus token rates for work that scores within a point of it.
Roughly 99% of Opus 4.6's SWE-bench Verified and OSWorld-Verified scores at 60% of the price, across a 1M-token context, with false refusals on benign prompts cut about 47x.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, Product Engineering Leads, Data Scientists, Security Engineers
- Deployment
- API-based, Cloud-only, Multi-cloud
- Founded
- 2021
- Headquarters
- San Francisco, United States
- Team Size
- 500+
Key Features
- ✓1M-token context window
Holds roughly 750,000 words in a single request at standard per-token rates across the entire window
- ✓Adaptive thinking
The model scales its own reasoning depth from contextual cues rather than requiring a manually configured thinking budget per call
- ✓Computer use at 72.5% OSWorld
An 11-point jump over Sonnet 4.5 that effectively matches Opus 4.6, making supervised GUI automation viable at Sonnet pricing
- ✓47x reduction in false refusals
Over-refusal on difficult benign prompts fell from 8.50% to 0.18%, unblocking security and moderation work earlier models stonewalled
- ✓128K output, 300K in batch
Emits up to 128,000 tokens synchronously and 300,000 through the Batch API extended-output beta header
- ✓Prompt caching
Cache reads bill at $0.30 per million tokens, a tenth of base input price, which dominates real cost in long agent loops
- ✓Prompt-injection hardening
Best-of-N browser attack success rate measured at 1.29% unmitigated and 0.51% with updated safeguards under ASL-3 deployment
Capabilities
Use Cases
- •Production coding agents
Powers IDE assistants and autonomous PR agents at 79.6% SWE-bench Verified for a fifth of frontier-model output cost
- •Browser and desktop automation
Drives supervised GUI workflows — form filling, data migration, legacy-app bridging — on the strength of its OSWorld-Verified score
- •Long-context document analysis
Reviews contracts, filings and specification sets inside the 1M-token window without chunking or retrieval-induced error
- •Security and moderation workflows
Handles legitimate-but-sensitive analysis that earlier Claude models refused at an 8.50% rate on benign prompts
- •Multi-step MCP and tool orchestration
Serves as the reasoning layer in Model Context Protocol agent stacks that call databases and internal APIs across many turns
Ideal For
Best For
- ✓Computer-use and browser automation agents — the 72.5% OSWorld-Verified score is an 11-point jump over Sonnet 4.5 and effectively ties Opus 4.6
- ✓Production coding assistants and IDE agents where 79.6% SWE-bench Verified at $3/$15 beats paying Opus rates for a point of difference
- ✓Security research, content moderation and other legitimate-but-sensitive domains that earlier Claude models over-refused at an 8.50% rate
- ✓Long-context agent loops that need to hold a whole repository or document set in the 1M-token window without a retrieval layer
- ✓Teams that want a frozen, pinned snapshot rather than tracking each new Anthropic release through their evaluation suite
Not Ideal For
- ✗Reasoning-heavy frontier evaluations — it scores 58.3% on ARC-AGI-2 against Opus 4.6's 68.8%, and BenchLM ranks it #44 of 128 on agentic capability, its weakest category
- ✗Unattended production computer use: 72.5% OSWorld still means roughly a quarter of attempts fail, so tool permissioning, domain allowlists and human approval for irreversible actions remain mandatory
- ✗Cost-dominated high-volume workloads — Haiku 4.5 at $1/$5 is a third of the price for classification and extraction that does not need this much capability
- ✗Anyone starting fresh today: Claude Sonnet 5 supersedes it and runs at $2/$10 introductory pricing through 31 August 2026, so 4.6 is currently both older and more expensive
- ✗Uncalibrated thinking budgets — reasoning tokens bill as output at $15 per million, and practitioners have reported 4.6's thinking duplicating itself and wasting tokens
Integrations
Deployment
Market Analysis
Pros
- ✓Best price-to-capability ratio in the 4.6 generation — within roughly a point of Opus 4.6 on both SWE-bench Verified and OSWorld-Verified at 60% of the token cost
- ✓The 47x drop in over-refusal on benign prompts (8.50% to 0.18%) materially widened what the model will actually attempt, with harmlessness on violative requests still rising to 99.38%
- ✓1M-token context at standard rates with 128K output, 300K through the Batch API
- ✓Anthropic's own blind preference testing put it ahead of Sonnet 4.5 about 70% of the time and ahead of the pricier Opus 4.5 59% of the time
Cons
- ✗Its own successor is cheaper — Claude Sonnet 5 runs at $2/$10 introductory pricing through 31 August 2026, so 4.6 is currently both older and more expensive for new work
- ✗Weakest on reasoning-heavy evaluations: 58.3% on ARC-AGI-2 against Opus 4.6's 68.8%, and BenchLM ranks it #44 of 128 on agentic capability, its lowest category placement
- ✗Not safe for unattended computer use despite the 72.5% OSWorld score — Caylent's production analysis insists on tool permissioning, domain allowlists, sandboxing and human approval for irreversible actions
- ✗Slow at about 44 output tokens per second on BenchLM's measurement against a 108 field median
- ✗The explicit extended-thinking parameter is deprecated on this model, and practitioners have reported its thinking output duplicating itself and burning tokens that bill at the output rate
- ✗Identity confusion has been reported in the wild — an HN-surfaced thread showed it claiming to be DeepSeek when prompted in Chinese
- ✗Reliability incidents are on the public record; Anthropic's status page logged a 'Claude Sonnet 4.6 Errors' incident
Pricing
Claude API (pay-as-you-go)
From $3/M input tokens
- ✓$3 per million input tokens
- ✓$15 per million output tokens
- ✓5-minute cache writes $3.75/M, cache reads $0.30/M
- ✓Full 1M context at standard rates
Batch API
From $1.50/M input tokens
- ✓50% discount on both input and output
- ✓$1.50/M input, $7.50/M output
- ✓Supports the 300K extended-output beta header
- ✓Stacks with prompt caching multipliers
Cloud marketplaces
Contact for pricing
- ✓Amazon Bedrock and Google Cloud billed by the provider
- ✓Claude Platform on AWS and Microsoft Foundry bill in Claude Consumption Units at $0.01 per CCU
- ✓US-only inference at a 1.1x multiplier via inference_geo
- ✓Global, multi-region and regional endpoint options
Token-metered at $3 input and $15 output per million with no seat licence, identical to Sonnet 4.5 — Anthropic held the price flat across the generation. The awkward commercial fact today is that its successor is cheaper: Claude Sonnet 5 runs at $2/$10 introductory pricing through 31 August 2026, reverting to the same $3/$15 afterwards, so through August 2026 Sonnet 4.6 is the more expensive of the two. Prompt caching drops reads to $0.30 per million and the Batch API halves both sides to $1.50/$7.50; the two stack. The full 1M context bills at standard rates with no long-context surcharge. Thinking tokens bill as output, which is where uncalibrated effort settings turn into real money. Server-side tools add on top — web search at $10 per 1,000 searches, web fetch free. US-only inference via `inference_geo` adds 1.1x; regional cloud endpoints add 10%.
Security & Compliance
Connect
Sources
This page was written from 9 sources, 8 on domains other than anthropic.com.
- 1.anthropic.com — claude sonnet 4 6vendor
- 2.platform.claude.com — pricing
- 3.platform.claude.com — model deprecations
- 4.platform.claude.com — overview
- 5.caylent.com — claude sonnet 4 6 in production capability safety and cost e
- 6.benchlm.ai — claude sonnet 4 6
- 7.openrouter.ai — claude sonnet 4.6
- 8.hn.algolia.com — hn.algolia.com
- 9.en.wikipedia.org — Anthropic
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.