Claude Haiku 5.5, Sonnet 5.5 and Mistral Large 4 are on Complete
New models on Complete
Three models landed in the last ten days, and one familiar model is on its way out. Claude Haiku 5.5 brings near flagship scores to high volume work at a few cents per task. Claude Sonnet 5.5 replaces Claude Sonnet 4.5, which Anthropic retires on November 30. Mistral Large 4 is in public preview, with open weights due by the end of October. All three are available for enterprise teams on Complete.
Every score below comes from the Artificial Analysis Intelligence Index, an independent rating built from 10 tests covering agent work, coding, general knowledge and science. Cost per task is what Artificial Analysis spent running each model through that test suite, divided by the number of tasks. Figures were captured on October 7, 2026.
Model | Released | Independent score | Cost per task | List price per million words in / out* |
|---|---|---|---|---|
Claude Haiku 5.5 (max effort) | Oct 7 | 43 | $0.21 | $0.10 / $0.50 |
Claude Sonnet 5.5 (max effort) | Sep 28 | 56 | $5.46 to $7.67** | $2 / $10 |
Mistral Large 4 preview | Oct 6 | 38 | $1.13 | $1.36 / $4.18 |
* Providers price per million tokens, which are fragments of words. Haiku 5.5 pricing applies to requests under 100,000 tokens.
** Artificial Analysis currently shows two figures for Sonnet 5.5 at max effort, $7.67 on its leaderboard and $5.46 on the model page.
Claude Haiku 5.5: above Sonnet 5.5 at medium effort, for a third of the cost
Haiku 5.5 is Anthropic's lowest priced current model, built for the steps that run thousands of times a day: routing requests, classifying tickets, pulling fields out of documents, drafting customer replies and running as a subtask worker inside larger AI workflows.
At its highest effort setting it scores 43 on the Artificial Analysis index for $0.21 per task. Claude Opus 5.5 scores 58 for $5.98. Haiku 5.5 also outscores Claude Sonnet 5.5 on its medium setting (41 at $0.59) for about a third of the cost.
Effort is the main dial. Lower settings cost far less and score lower:
Effort setting | Score | Cost per task |
|---|---|---|
Max | 43 | $0.21 |
Extra high | 41 | $0.12 |
High | 38 | $0.08 |
Medium (default) | 34 | $0.05 |
Low | 29 | $0.02 |
What else to know:
Reads text and images, writes text. It holds up to 1 million tokens of context in a single request.
Pricing rises to $0.50 in and $2.50 out for requests over 100,000 tokens. Batch jobs cost half.
Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5. HubSpot reports 92.8% on its internal CRM test suite, and Asana reports responses more than 30% faster. These are vendor and customer figures, not independent ones.
Haiku 5.5 thinks at length before it answers. Artificial Analysis counted more output than Opus 5.5 at max effort, so the effort setting has a large effect on real bills.
Anthropic still points customers to Sonnet and Opus for complex coding work.
It is not the cheapest route to a score in the mid 40s. MiMo-V2.6-Pro, an open model from Xiaomi, scores 46 for $0.13 per task on the same index.
Two changes matter for anyone moving from Haiku 4.5. A new tokenizer counts the same text as roughly 30% more tokens, and the temperature, top_p and top_k settings now return an error.
Claude Sonnet 5.5 replaces Sonnet 4.5 on November 30
Anthropic retires Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) on November 30, 2026, and names Claude Sonnet 5.5 (claude-sonnet-5-5) as the replacement. Any workflow still pinned to the Sonnet 4.5 model ID stops working after that date. Amazon Bedrock and Google Cloud Vertex AI set their own retirement dates.
The replacement costs less and runs faster:
List price drops from $3 in and $15 out to $2 in and $10 out, a third lower.
Artificial Analysis measures Sonnet 5.5 at 141 tokens per second against 43 for Sonnet 4.5, over 3x faster.
Context grows to 1 million tokens, and Anthropic has cut the price of cached input to $0.10 per million tokens.
The effort setting decides most of the cost. On the Artificial Analysis leaderboard, cost per task rises about 18x from the lowest setting to the highest:
Effort setting | Score | Cost per task |
|---|---|---|
Max | 56 | $7.67 |
Extra high | 52 | $2.75 |
High | 47 | $1.12 |
Medium | 41 | $0.59 |
Low | 36 | $0.42 |
For a migration, three decisions come first: which effort setting each workflow needs, whether high volume steps should move down to Haiku 5.5 instead, and when to switch before the November 30 cutoff. Artificial Analysis does not publish a cost per task for Sonnet 4.5, so a like for like cost comparison is not available.
Mistral Large 4 preview: open weights on the way
Mistral Large 4 is Mistral's new flagship, released as a public preview on October 6. It reads documents, charts and images alongside text. Mistral says the open weights will ship by the end of October, which will let companies run it on their own infrastructure. Mistral has not yet named the license.
Scores 38 on the Artificial Analysis index at $1.13 per task.
Runs at 116 tokens per second on Artificial Analysis, close to Sonnet 5.5.
Writes far shorter answers than Haiku 5.5 or Sonnet 5.5 at max effort, which keeps output costs down.
List price is $1.36 in and $4.18 out per million tokens. Artificial Analysis notes a promotional price during the first two weeks, so early prices on some providers sit below list.
Mistral reports 61.7% on DeepSWE, a software engineering test, and 59.9% on AutomationBench, a business automation test. These are Mistral's own results.
On today's independent numbers, Gemini 3.8 Flash and Ling 3.1 Flash score 41 at a similar cost. The case for Mistral Large 4 rests on the open weights and on a European provider. Independent scores and prices will move once the weights ship and the promotional pricing ends.
Run them on your own workflows
Claude Haiku 5.5, Claude Sonnet 5.5 and Mistral Large 4 preview are available for enterprise teams on Complete. Book a demo to run them against your own workflows and plan your Sonnet 4.5 migration before November 30.
Read the full post on each model:
Claude Sonnet 5.5 and the Sonnet 4.5 retirement
