Anthropic

Claude Sonnet 4.6

Type
language model
Context
1,000K tokens
Max output
64K
Released
17 February 2026
Knowledge cutoff
August 2025
API string
anthropic/claude-sonnet-4-6

Prices by provider

Provider Input, $ per 1M Output, $ per 1M In our data since
xpersona $0.9 $5.55 31 Jul 2026
Kilo Code stealth/claude-sonnet-4.6 $2.4 $12 31 Jul 2026
Poe $2.6 $13 31 Jul 2026
NanoGPT $2.992 $14.994 31 Jul 2026
daoxe $3 $15 31 Jul 2026
neon $3 $15 31 Jul 2026
frogbot $3 $15 31 Jul 2026
auriko $3 $15 31 Jul 2026
google-vertex-anthropic $3 $15 31 Jul 2026
routing-run $3 $15 31 Jul 2026
freemodel $3 $15 31 Jul 2026
perplexity-agent $3 $15 31 Jul 2026
modelis $3 $15 2 Aug 2026
302.AI $3 $15 31 Jul 2026
AIHubMix $3 $15 31 Jul 2026
Abacus.AI $3 $15 31 Jul 2026
Amazon Bedrock anthropic.claude-sonnet-4-6 $3 $15 31 Jul 2026
Anthropic $3 $15 31 Jul 2026
Cloudflare AI Gateway $3 $15 31 Jul 2026
CrossModel $3 $15 31 Jul 2026
Databricks $3 $15 31 Jul 2026
DigitalOcean Gradient AI $3 $15 1 Aug 2026
FastRouter $3 $15 31 Jul 2026
GMI Cloud $3 $15 31 Jul 2026
GitHub Copilot $3 $15 31 Jul 2026
Google Vertex AI $3 $15 31 Jul 2026
Kilo Code anthropic/claude-sonnet-4.6 $3 $15 31 Jul 2026
LLM Gateway $3 $15 31 Jul 2026
Merge Gateway $3 $15 31 Jul 2026
Microsoft Foundry $3 $15 31 Jul 2026
NEAR AI $3 $15 31 Jul 2026
OpenCode Zen $3 $15 31 Jul 2026
OpenRouter $3 $15 31 Jul 2026
OrcaRouter $3 $15 31 Jul 2026
Pioneer $3 $15 31 Jul 2026
Requesty $3 $15 31 Jul 2026
SAP AI Core $3 $15 1 Aug 2026
Snowflake $3 $15 31 Jul 2026
Vercel $3 $15 31 Jul 2026
ZenMux $3 $15 31 Jul 2026
Amazon Bedrock au.anthropic.claude-sonnet-4-6 $3.3 $16.5 31 Jul 2026
Cortecs $3.59 $17.92 1 Aug 2026
Venice AI $3.6 $18 31 Jul 2026

Only the base tier and only per-token prices are shown. Batch, discounted and cached rates, as well as prices per image or per second of video, do not go into this table: they cannot stand in the same column as a price per million tokens.

A single provider appears several times if it sells this model under different identifiers and at different prices — for example, at different weight precision. The identifier is shown next to the name. Identical offers that arrived from two sources under different spellings are merged into one row.

The date in the last column is the day this price first entered our collection. The price may well be older: before that day we simply were not recording it. It has not changed since — otherwise a new row with a new date would stand in its place.

Measurement results

Task set Result Run conditions Measured by
Arena Score, web development 1,522.64 1,517–1,528
Arena Score, programming 1,502.48 1,497–1,508
Arena Score, expert questions 1,501.16 1,492–1,510
Arena Score, hard prompts 1,484.55 1,480–1,489
Arena Score, long queries 1,480.85 1,475–1,486
Arena Score, instruction following 1,477.49 1,472–1,483
Arena Score in English 1,473.19 1,468–1,478
Arena Score in French 1,464.7 1,449–1,481
Arena Score, multi-turn dialogue 1,463.39 1,457–1,470
Arena Score in Spanish 1,462.54 1,447–1,478
Arena Score, mathematics 1,461.43 1,450–1,472
Arena Score, overall 1,457.32 1,454–1,461
Arena Score in Russian 1,438.4 1,430–1,447
Arena Score, creative writing 1,438.07 1,431–1,445
Arena Score, understanding diagrams 1,306.39 1,297–1,315
Arena Score, text recognition in images 1,295.17 1,289–1,302
Arena Score, working with images 1,281.47 1,275–1,288
ARC-AGI — generalising to unseen patterns 86.5 % effort: high · with a tuned harness
Mock AIME 2024–2025 — olympiad problems 85.78 % effort: high · with a tuned harness Epoch evaluations
GPQA Diamond — graduate-level questions 83.17 % effort: high · with a tuned harness Epoch evaluations
SWE-bench Verified — fixing bugs in repositories 75.21 % · with a tuned harness Epoch evaluations
OSWorld — working inside an operating system, first version 72.1 % · with a tuned harness OS World Website
WeirdML — unusual machine learning tasks 66.07 % effort: medium WeirdML Leaderboard
ARC-AGI-2 60.42 % effort: high
DeepResearch Bench — deep research 54.87 % effort: high https://drb.futuresearch.ai/#drb self-reported
Terminal-Bench — working in the command line 53.4 % · with a tuned harness https://www.tbench.ai/leaderboard/terminal-bench/2.0
CursorBench — edits in the editor 49 % effort: max https://cursor.com/cursorbench
SimpleQA Verified — factual accuracy 29 % thinking budget: 32K Epoch evaluations
APEX-Agents 23.7 % effort: high
OSWorld 2.0 — working inside an operating system 9.3 % effort: medium https://osworld-v2.xlang.ai/ self-reported
Chess puzzles 8.46 % effort: high Epoch evaluations
CritPt — physics problems 3.14 % effort: max
Arena Score, agent tasks 0.03 0–0