New Sign up free, 10 calls on us. Up to $1, no card needed.

Claude Sonnet 4.6

Released 2026-02-17

chatCodeReasoningTool callingVisionPrompt caching

Claude Sonnet 4.6 is the generation that brought big-context capability to the Sonnet tier: versus Sonnet 4.5 it expands the context window from 200K to 1M tokens, doubles max output to 128K tokens, and adds adaptive thinking alongside extended thinking.

Input
text image $3/M
Output
text $15/M
Cache read
$0.3/M
Context
1M
vs GPT-4o
~40% cheaper
Knowledge cutoff
2025-08

Benchmarks

Above averageonly 2 comparable
Claude Sonnet 4.6 other models measured peer average no peer scored higher
Terminal Bench 2.0 Terminus 2, max
59.1%
OSWorld-Verified
72.5%
GDPval-AA v2 Elo, max · 1606-1633
1633
MMMLU max
89.3%
MCP-Atlas max
61.3%
MMMU-Pro no tools, max
74.5%

Vendor-published: Alibaba (Qwen) Anthropic ByteDance Google Moonshot OpenAI

Price in context

Where the price sits among 60 comparable models

Input$3/M
$0.05 · Qwen3 VL Flash GPT-5.4 Pro · $30
Output$15/M
$0.275 · DeepSeek V4 Flash GPT-5.4 Pro · $180
Cached read$0.3/M
$0.0028 · DeepSeek V4 Flash GPT-5.4 Pro · $15

The bar shows how this model’s price compares with every other model of the same kind on Synthorai. The cheapest and the most expensive are named at each end. These are base rates; batch, region and cache-write discounts are on the pricing page.

Specs & limits

Tokens

Context window (vendor spec) 1,000,000
Max output (vendor spec) 128,000
Knowledge cutoff 2025-08
Training data through 2026-01 knowledge cutoff 2025-08

Prompt caching

How it caches explicit (opt-in)
Min prefix 1,024
Lifetime 5m default, 1h option
Write cost 1.25x (5m) / 2x (1h)

Thinking

Vendor control thinking.type + output_config.effort (budget_tokens still accepted, deprecated)
Accepted values thinking.type adaptive · enabled (deprecated) · disabled; effort low · medium · high · max (no xhigh)
Default thinking off; effort high applied when the request sets nothing
Can be turned off Yes
Thinking behaviour Adaptive thinking interleaves between tool calls automatically with no beta header; in manual "enabled" mode the interleaved-thinking-2025-05-14 header still works but is deprecated; earlier turns' thinking blocks are kept in context and billed as input.
Parameter reasoning_effort
Values minimal · low · medium · high the gateway's parameter surface - the vendor mapping above applies

Model

Modalities text + image → text
  • 1M-token context GA at standard pricing since 2026-03-13
  • up to 300k output tokens on Batch API with output-300k-2026-03-24 beta header

per Anthropic official docs ↗

One prompt, measured through the gateway

PROMPT Explain to a CFO, in exactly three sentences, why our LLM bill doubled after we enabled reasoning. CHECK exactly 3 sentences

Claude Sonnet 4.6 passed · 3 sentences

Here is a three-sentence explanation for a CFO: Reasoning models "think out loud" by generating a long internal chain of thought before producing the final answer, and every token in that thinking process is billed at the same rate as the output you actually see. This means a single user request that previously cost 500 tokens of output may now cost 5,000 tokens once the hidden reasoning steps are counted, multiplying cost per call by roughly 5-10x depending on task complexity. To control this, we can cap reasoning depth, route simpler tasks back to non-reasoning models, and audit which use cases actually require that level of accuracy to ensure we are only paying the premium where it delivers measurable business value.

out 156 tok latency 4.9 s

Instruction following (exactly three sentences - countable), audience shaping (CFO register), and the hidden-thinking billing gap the token meters below expose.

PROMPT This function is supposed to return the longest run of consecutive equal items, but callers report it is off by one on single-element inputs and misses runs that end at the last index. Fix it and explain the bug in one paragraph. CHECK fix passes tests

Claude Sonnet 4.6 passed · 8/8 cases

## The Bug There are two related off-by-one errors: 1. **`cur` starts at `0` instead of `1`** — a single matching pair increments `cur` to `1`, which correctly represents a run of 2, but a single-element list (or any non-repeating list) leaves `cur = 0` and `best = 0`, when the answer should be `1`. 2. **`best` is only updated inside the `if` branch** — a run that ends at the last index is captured correctly *in that branch*, but the initialization problem means the baseline "every element is a run of at least 1" is never established. The root cause is that **`cur` tracks the number of *consecutive equal pairs*, not the length of the run**. Switching to tracking run length directly (starting `cur = 1`, updating `best` on *every* iteration including the `else` branch reset) fixes both issues. […]

out 685 tok latency 12.2 s

Whether the fix is actually correct (runnable), explanation density, and token efficiency on a bounded task.

PROMPT Extract a JSON object with fields {company, quarter, revenue_usd_m, growth_pct, guidance} from this text. Use null for anything not stated; add no other fields. CHECK valid JSON, schema exact

Claude Sonnet 4.6 passed · 5/5 fields, guidance null

```json { "company": "Meridian Robotics", "quarter": "Q3 (September quarter)", "revenue_usd_m": 412, "growth_pct": 11, "guidance": null } ```

out 67 tok latency 1.3 s

Schema obedience (no invented fields), hallucination pressure (guidance is explicitly withheld), and structured-output path differences.

PROMPT Write a 120-word product announcement for an API gateway feature that caches prompts across providers. Forbidden words: "seamless", "unlock", "game-changer", "revolutionize", "empower". Exactly one sentence must be a question. CHECK 120 words, 0 banned words

Claude Sonnet 4.6 passed · 121 words, 0 banned, 1 question

**Introducing Prompt Caching Across Providers** Managing AI costs gets complicated when identical prompts hit multiple providers repeatedly, burning through tokens unnecessarily. Today, we're launching cross-provider prompt caching in our API gateway — a feature that stores prompt responses and serves cached results regardless of which underlying model handles your request. Here's how it works: when your application sends a repeated prompt, the gateway checks its cache first, returning stored outputs in milliseconds rather than forwarding requests to OpenAI, Anthropic, or any connected provider. Why pay for the same generation twice? With configurable TTL settings and cache-key customization, teams processing high-volume or repetitive workloads will see immediate latency improvements and measurable cost reductions starting on day one. Available now on all paid plans.

out 177 tok latency 5.4 s

Constraint obedience (word budget, banned-word list, the single question), style fingerprint, and length control.

Use Claude Sonnet 4.6 in 30 seconds

OpenAI-compatible: swap the base_url, keep your SDK. POST /v1/chat/completions

from openai import OpenAI

client = OpenAI(
    base_url="https://synthorai.io/v1",
    api_key="sk-syn-...",
)

resp = client.chat.completions.create(
    model="claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Summarize this diff"}],
    reasoning_effort="medium",
)
print(resp.choices[0].message.content)

About Claude Sonnet 4.6

  • Anthropic launched it as its balanced model combining speed and intelligence for everyday tasks, with improved agentic search performance while consuming fewer tokens, and the 1M window is generally available at standard pricing rather than beta-gated.
  • It keeps Sonnet's fast latency profile and $3/$15-per-million-token pricing, with vision input and tool use throughout, and it qualifies for the Batch API's extended 300K-output beta.
  • Effort covers low through max but not xhigh; although the parameter defaults to high, Anthropic explicitly recommends setting it on this model to avoid unexpected latency, and suggests medium as the practical default.
  • Extended thinking still works but is deprecated in favour of adaptive, and neither thinking type is rejected outright.
  • Prefilling the assistant message returns an error here, while non-default sampling parameters are still tolerated.
  • That restriction arrives with Sonnet 5.
  • It uses the older tokenizer, so token counts stay comparable with earlier models.
  • Anthropic now lists it as a legacy model superseded by Claude Sonnet 5.
  • Through Synthorai's OpenAI-compatible endpoint it drops into any existing GPT-style integration unchanged.

FAQ

Is the Claude Sonnet 4.6 API free to try?

Yes: new accounts get 10 trial calls and up to $1 in free credit, no card required. At $3/M input tokens, that credit alone covers roughly 41 requests of ~8K tokens against Claude Sonnet 4.6.

What is Claude Sonnet 4.6 best at?

Context expanded from 200K to 1M tokens; doubled max output to 128K tokens; adaptive thinking at unchanged $3/$15 pricing. See the About section for the full picture from the vendor's own release notes.

How much does Claude Sonnet 4.6 cost?

Claude Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens on Synthorai. That is the provider's list price, with no platform markup. Cached input tokens bill at $0.3/M.

Does Claude Sonnet 4.6 support prompt caching?

Yes, via opt-in: mark stable prefixes with cache_control breakpoints. Cached input tokens bill at $0.3/M vs $3/M uncached; prompts need a 1,024-token stable prefix to cache (TTL 5m default, 1h option). Prompt caching guide →

How do I get access to Claude Sonnet 4.6?

Point your existing OpenAI SDK at base_url="https://synthorai.io/v1", set model="claude-sonnet-4-6", and you're done. One API key covers every model on the gateway.

What is Claude Sonnet 4.6's knowledge cutoff?

Claude Sonnet 4.6's knowledge cutoff is 2025-08, per the vendor's official documentation (as of 2026-07-09).

Related models

Compare

Every value on this page is transcribed from the vendor's own documentation, linked above, and carries the date it was checked. Prices are compared across the catalogue; specification values that vendors define differently are shown with the difference stated rather than charted. Nothing here is measured by us, and nothing is scored.

Get your API key Compare your cost →