Best Model for Continue

Continue is no longer maintained: its team joined Cursor and the repo has been read-only since mid-2026, so the extension won't learn new models or API features, and Cline or Kilo Code is the maintained pick for a new setup. For an existing install, Claude Sonnet 5.5 is the best model for Continue's agent mode: it scores 61.8% on Terminal-Bench 4.0 at $2/$10 per million tokens, half the price of Claude Opus 5.5 (64.9%). Continue assigns models by role (chat, edit, apply, autocomplete, embedding and reranking), so the agent model and the autocomplete model are separate choices, and a small, fast model such as Codestral suits autocomplete. Continue's docs warn that small local models have limited tool calling, which makes agent mode hard, so local setups work best for autocomplete and chat.

Read the full Continue review → · updated 2026-10-17

Sources: Continue model docs · Continue repo (read-only) · Terminal-Bench 4.0 leaderboard · Anthropic on Haiku 5.5 · DeepSeek V4.1 Flash release

#1Claude Sonnet 5.5Anthropic · $4.40/1M · 1M ctx

Best overall. The agent and chat model: 61.8% on Terminal-Bench 4.0 and 95.8 on Scale's SWE-Bench Pro V2, with a 1M-token context.

Overall 96/100 · Value 59Full breakdown →

#2Claude Opus 5.5Anthropic · $8.80/1M · 1M ctx

Best for long agent runs. #1 on Terminal-Bench 4.0 at 64.9%, for long agent-mode tasks where a failed run costs more than the tokens.

Overall 98/100 · Value 51Full breakdown →

#3Claude Haiku 5.5Anthropic · $0.22/1M · 1M ctx

Best value. A cheap chat and edit model at $0.10/$0.50 per million tokens; Anthropic reports 39.2% on Terminal-Bench 4.0. Prompts over 100K tokens cost five times as much.

Overall 82/100 · Value 97Full breakdown →

#4CodestralMistral · $0.48/1M · 256K ctx

Autocomplete pick. Mistral's code model built for fill-in-the-middle completions, at $0.30/$0.90 per million tokens with a 256K context. It is not meant for agent mode.

Overall 53/100 · Value 58 · 64% community winFull breakdown →

#5DeepSeek V4.1 FlashDeepSeek · $0.22/1M · 1.0M ctx

Budget pick. Open weights under the MIT license at $0.30/$1.20 per million tokens at peak hours and half off-peak; DeepSeek reports 42.3% on Terminal-Bench 4.0 for its October 17 update, up from 31.2%.

Overall 77/100 · Value 93Full breakdown →

#6GPT-6.1 SolOpenAI · $4.40/1M · 1.1M ctx

OpenAI pick. 58.2% on Terminal-Bench 4.0 at the same $2/$10 list price as Sonnet 5.5.

Overall 92/100 · Value 58Full breakdown →

FAQ

What is the best model for Continue?

Claude Sonnet 5.5 for agent mode and chat (61.8% on Terminal-Bench 4.0), Claude Opus 5.5 for long, hard agent tasks, and a small fast model such as Codestral for autocomplete.

Can Continue run fully local models?

Yes, through Ollama and similar runtimes. Continue's docs list Qwen3 Coder 30B, Qwen2.5-Coder 7B and Devstral Small 27B for local use and warn that their limited tool calling makes agent mode hard, so many setups pair a local autocomplete model with a hosted agent model.

Which model should I use for Continue's autocomplete?

A small, fast code model. Continue's docs suggest Mercury Coder among hosted options and QwenCoder2.5 1.5B or 7B locally; Codestral is a low-cost hosted alternative at $0.30/$0.90 per million tokens.

Is Claude Sonnet 4.6 still the best hosted model for Continue?

Claude Sonnet 5.5 has replaced it at a lower list price ($2/$10 per million tokens against $3/$15).

Is Continue still maintained?

No. The Continue team joined Cursor and the continuedev/continue repo is read-only, so there are no fixes for new models or editor versions. The extension still installs and runs; for a maintained open-source extension with the same model freedom, use Cline or Kilo Code.

Current Claude, GPT, Gemini and open models run on Standard Compute

Standard Compute

Current frontier models Claude Fable 5.1 · GPT-6 Astra