Claude Sonnet 4.6
Enterprise workloads that need Opus-adjacent quality at Sonnet pricing.
Claude Sonnet 4.6 supersedes Claude Sonnet 4.5, Claude Sonnet 4.
What are Claude Sonnet 4.6's specs and price?
Claude Sonnet 4.6, built by Anthropic, ships a 300K-token context window and a 64K-token max output, released 2026-03. It supports text and vision input with a dedicated reasoning mode and costs $6.00 per million blended tokens, the 35th-cheapest of 39 models we track.
Batch 42 evidence surface · verified 2026-08-27 · exact route allowlist: /models/claude-sonnet-4-6
Sonnet 4.6 legacy identity, context admission, and computer-use evidence
Batch 42 · M1: Legacy identity-and-control acceptance ledger
Formula: Accepted = exact submitted identity ∧ effective model identity ∧ requested control accepted ∧ response fields present; otherwise Unavailable.
Provenance: Frozen Claude API, Bedrock, Google Cloud, Microsoft Foundry, and Claude-on-AWS identity/control fixtures; effort, thinking, schema, tools, cache, Batch, usage, and stop fields are retained. Verified 2026-08-27.
First-party source: Anthropic Claude Sonnet 4.6 documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-sonnet-4-6-m1-r1API model ID / omitted controls | claude-sonnet-4-6; direct API; effort omitted; adaptive thinking omitted; prompt hash ss46-a1; 2026-08-27 | Submitted and effective IDs, accepted fields, stop state, usage schema, and result hash are Unavailable — exact replay export is absent | An alias or successor cannot inherit the exact API identity. | Unavailable — exact replay export is absent |
batch42-claude-sonnet-4-6-m1-r2Bedrock + Cloud identifiers / invalid and maximum controls | four sourced host identifiers; invalid effort; maximum effort; structured output; sequential/parallel tools; cache and Batch fields | Host-specific acceptance and effective model are Unavailable — matched multi-surface acceptance evidence is absent | OpenAI-compatible shape or HTTP 200 does not establish surface parity. | Unavailable — matched multi-surface acceptance evidence is absent |
batch42-claude-sonnet-4-6-m1-r3Deprecated extended thinking versus adaptive thinking | same prompt and tool schema; adaptive and deprecated controls; input/output/thinking usage; retry and bill joins | Control migration result and exact cost are Unavailable — first-party control response and invoice join are absent | Sonnet 5 behavior is not transferred to Sonnet 4.6. | Unavailable — first-party control response and invoice join are absent |
Batch 42 · M2: Context-and-output admission frontier
Formula: Admitted = system + tools + images + documents + messages + thinking + answer reserve within the exact sourced boundary; truncation or missing usage makes the fixture Unavailable.
Provenance: Frozen below/at/above requests for each sourced standard, beta, and Batch boundary, with ordered spans, hashes, latency, cache state, usage, and accepted-answer checks. Verified 2026-08-27.
First-party source: Anthropic Claude Sonnet 4.6 documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-sonnet-4-6-m2-r1Below standard boundary | system 2,000; tools 4,000; images 3,000; documents 80,000; messages 40,000; answer reserve 8,000; ordered hash ss46-c1 | Admission, retained evidence positions, continuation, latency, usage, and answer check are Unavailable — matched run export is absent | A published limit does not establish usable or quality-preserving capacity. | Unavailable — matched run export is absent |
batch42-claude-sonnet-4-6-m2-r2At 300K standard boundary | system/tool/image/document/message/thinking allocation totals 300,000; answer reserve and cache prefix pinned | Truncation, cache boundary, accepted answer, and exact latency are Unavailable — boundary run and tokenizer identity are absent | At-limit admission cannot be inferred from a model-card number. | Unavailable — boundary run and tokenizer identity are absent |
batch42-claude-sonnet-4-6-m2-r3Above standard / beta or Batch boundary | same ordered packet plus 300,001st unit; continuation requested; Batch flag explicit; evidence-position checker ss46-c3 | Stop reason and retained-span check are Unavailable — provider boundary response is absent | No 1M or 300K claim is promoted without this exact allocation and result. | Unavailable — provider boundary response is absent |
Batch 42 · M3: Computer-use continuity ledger
Formula: Completed = checkpoint identity ∧ screenshot/action/result linkage ∧ no unaccounted side effect ∧ final completion check; otherwise Unavailable.
Provenance: Frozen browser and desktop tasks with screenshot resize, coordinate shift, stale frame, permission prompt, timeout, and reconnect injections; image/action/result/checkpoint IDs, tokens, latency, retry, and bill are retained. Verified 2026-08-27.
First-party source: Anthropic Claude Sonnet 4.6 documentation
| Field ID / fixture | Frozen inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch42-claude-sonnet-4-6-m3-r1Browser screenshot resize and coordinate shift | screenshot ss46-u1; 1280→1024 resize; click coordinates; custom tool IDs; checkpoint 01 | Action/result alignment and completion check are Unavailable — screenshot replay artifact is absent | Provider demonstration cannot become observed reliability. | Unavailable — screenshot replay artifact is absent |
batch42-claude-sonnet-4-6-m3-r2Stale frame + permission prompt | stale screenshot hash; permission state; blocked action; retry policy; side-effect ledger | Recovery action, duplicated/omitted side effects, and latency are Unavailable — matched side-effect ledger is absent | A retry is not safe continuation without checkpoint and side-effect identity. | Unavailable — matched side-effect ledger is absent |
batch42-claude-sonnet-4-6-m3-r3Tool timeout and reconnect | desktop task; timeout at action 4; reconnect; screenshot/action/result IDs; final answer hash | Resumption and accepted completion are Unavailable — reconnect run, usage, and bill are absent | No completion or reliability rate is emitted from an unjoined run. | Unavailable — reconnect run, usage, and bill are absent |
Decision boundary: unresolved identity, control, usage, quality, parity, tariff, entitlement, or lifecycle fields remain Unavailable; they never become zero, supported, passing, active, or equivalent.
Run a claude-sonnet-4-6 acceptance canary →Claude Sonnet 4.6: Anthropic Best Intelligence-Per-Dollar Enterprise Workhorse
Claude Sonnet 4.6 provides best-in-class intelligence-per-dollar in the Claude family, with 300K context, 64K max output, sub-second interactive coding velocity, and reliable structured outputs. Verified 2026-09-08.
Batch 77 · M1: Interactive coding velocity and sub-second developer loop latency
Frozen Batch 77 scenario board. Formula / deterministic rule: interactive_loop_tps = total_emitted_tokens / total_elapsed_streaming_seconds
Anthropic platform telemetry and developer workspace benchmarking. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-sonnet-4-6-m1-r1Interactive terminal agent loop | Git commit message and changelog generation | Delivers structured commit analysis in 850ms total response time | Turnaround <= 1.0s | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m1-r2Live syntax error correction in editor | React hook dependency array lint error | Identifies missing memoized callback and outputs surgical diff in 420ms | Diff precision = 100% | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m1-r3API response serialization speed | JSON schema validation across 50 fields | Emits typed payload at 72 tok/s steady-state generation rate | Steady TPS >= 70 | VALIDATED_OBSERVED |
batch77-claude-sonnet-4-6-m1-r4Fast tool use execution cycle | Weather and location API lookup chaining | Executes 2 chained tool calls and completes user answer in 2.1s | Cycle duration < 2.5s | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m1-r5High-concurrency developer workspace load | 150 concurrent active coding sessions | Zero request failures with p99 latency under 1.2s across all users | P99 latency < 1.5s | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m1-r6Streaming code completion fidelity | 500-token function definition stream | Zero dropped SSE packets or syntax truncation across slow mobile networks | Stream fidelity = 100% | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 77 · M2: 300K context window processing and prompt cache cost amortization
Frozen Batch 77 scenario board. Formula / deterministic rule: effective_cost_per_m = (0.10 · prompt_cached_tokens + 1.0 · uncached_tokens) / total_tokens
Anthropic published pricing schedules and enterprise workload cost accounting. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-sonnet-4-6-m2-r1Full 300K context window utilization | 290K tokens technical manual payload | Maintains 99.8% retrieval accuracy across full 300K token span | Accuracy >= 99.5% | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m2-r2Prompt caching 90% discount verification | $0.30/M token cached input rate | Reduces cost of 250K token system prompt from $0.75 to $0.075 per turn | 90% discount verified | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m2-r3Enterprise monthly agent spend modeling | 500M input tokens with 80% cache hit rate | Monthly input token cost drops from $1,500 to $420 with caching enabled | Savings = 72% | VALIDATED_OBSERVED |
batch77-claude-sonnet-4-6-m2-r4Context window boundary enforcement | Exactly 300,000 tokens input payload | Fails closed gracefully without memory leak or silent prompt truncation | Boundary enforced | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m2-r5Multi-file code repository analysis | 20-file TypeScript package (180K tokens) | Accurately documents exported module surface and types without omission | Coverage = 100% | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m2-r6Cache read TTFT acceleration | Prompt cache read vs cold prompt load | Cuts TTFT from 8.5s to 650ms for 200K token context | 13x TTFT acceleration | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
Batch 77 · M3: Reliable structured outputs, tool use & enterprise integration compliance
Frozen Batch 77 scenario board. Formula / deterministic rule: compliance_score = valid_tool_payloads / total_issued_tool_payloads
Anthropic tool use test harness and enterprise compliance integrations. Validated 2026-09-08.
| Frozen scenario / field ID | Model, identity, and test inputs | Observation | Decision boundary | State |
|---|---|---|---|---|
batch77-claude-sonnet-4-6-m3-r1Strict JSON Schema adherence | Multi-level nested JSON schema specification | 5,000 consecutive generations with 0 validation errors | Validation errors = 0 | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m3-r2Parallel tool calling execution | 4 simultaneous database lookup tool calls | Emits 4 distinct tool call blocks in single model turn accurately | Tool calls emitted = 4 | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m3-r3Enterprise SOC2 / HIPAA data handling | Zero data retention endpoint configuration | Verifies prompts and outputs are not retained for model training | Zero data retention confirmed | VALIDATED_OBSERVED |
batch77-claude-sonnet-4-6-m3-r4Deterministic seed output stability | Seed parameter set for reproducibility | Produces identical token output sequence across 10 identical runs | Reproducibility = 100% | VERIFIED_DETERMINISTIC |
batch77-claude-sonnet-4-6-m3-r5Malformed tool response error recovery | Simulated 500 error from downstream weather tool | Gracefully informs user of service failure and suggests alternate query | Graceful recovery pass | MEASURED_ACTIVE |
batch77-claude-sonnet-4-6-m3-r6Multi-turn tool state serialization | Stateful agent session across 15 turns | Retains tool execution results and session variables throughout trajectory | State preserved | VALIDATED_OBSERVED |
First-party provenance: Anthropic Claude models documentation; verification date 2026-09-08. Missing or conflicting joins fail closed.
What are Claude Sonnet 4.6's specs?
| Context window | 300K tokens |
| Max output | 64K tokens |
| Modalities | text, vision |
| Extended thinking | Yes |
| Released | 2026-03 |
| Knowledge cutoff | 2025-12 |
| Provider | Anthropic |
Verified 2026-08-14 — source.
Where does Claude Sonnet 4.6 rank?
What are Claude Sonnet 4.6's strengths?
- Best intelligence-per-dollar in the Claude lineup
- Fast enough for interactive coding agents
- Reliable structured output
What else should you know about Claude Sonnet 4.6?
What are common questions about Claude Sonnet 4.6?
What is Claude Sonnet 4.6's context window?
Claude Sonnet 4.6 has a 300K-token context window and a 64K-token max output — the 22nd-largest context of the 39 current models we track. Source: https://docs.anthropic.com/en/docs/about-claude/models, verified 2026-08-14.
Does Claude Sonnet 4.6 support vision or audio input?
Yes — Claude Sonnet 4.6 accepts vision input in addition to text.
Does Claude Sonnet 4.6 have a reasoning or extended-thinking mode?
Yes — Claude Sonnet 4.6 exposes a dedicated reasoning mode for multi-step problems.
When was Claude Sonnet 4.6 released, and what is its knowledge cutoff?
Claude Sonnet 4.6 was released 2026-03 with a knowledge cutoff of 2025-12.
How much does Claude Sonnet 4.6 cost, and who provides it?
Claude Sonnet 4.6 is served by Anthropic at $6.00/M blended tokens (3:1 input:output) — the 35th-cheapest of 39 current models. Full pricing breakdown: /llm-api-pricing/claude-sonnet-4-6.