On September 28, Cognition announced that Claude Sonnet 5.5 is now available in Devin Desktop and Devin CLI, delivering 64.4% on FrontierCode 1.1 Main — a 14.6-point improvement over Sonnet 5.0’s 56.2% baseline. The company also noted that Sonnet 5.5 surpasses Fable 5.1 when running at extra-high reasoning effort, making it the highest-scoring model available in Devin’s local and CLI environments.
The update closes the model availability gap that existed between Devin Cloud (which got Sonnet 5.5 in August) and Desktop/CLI users, who had been waiting for Anthropic to release the model via its API. Now all three Devin surfaces run the same frontier reasoning models.
What changed: 64.4% vs 56.2%
FrontierCode 1.1 Main is Cognition’s internal benchmark for measuring how often an AI agent successfully completes a real-world coding task (bug fix, feature implementation, or refactor) on the first try, with the resulting code passing tests and ready to merge. Sonnet 5.0 scored 56.2% on this benchmark when Cognition first added it to Desktop and CLI in June 2026. Sonnet 5.5 scores 64.4% — a 14.6-point gain — meaning it completes nearly 15% more tasks without needing a second attempt or human intervention.
The jump comes from improvements in reasoning depth and long-context accuracy. Anthropic’s Sonnet 5.5 release notes (published August 28, 2026) highlight two technical changes:
- Extended context window: Sonnet 5.5 supports up to 500,000 tokens of context (up from 200,000 in Sonnet 5.0), letting it hold more of a codebase in memory during multi-file refactors.
- Improved reasoning effort scaling: Sonnet 5.5 can allocate more compute to hard problems by running in “extra-high reasoning” mode, which takes longer but produces better results on tasks that require multi-step planning.
That second point is where Sonnet 5.5 pulls ahead of Fable 5.1. Cognition’s announcement notes that Sonnet 5.5 at extra-high reasoning surpasses Fable 5.1 on FrontierCode, making it the best-performing model available in Desktop and CLI when quality matters more than speed.
What “surpassing Fable 5.1 at extra-high reasoning” means
Fable 5.1 is Anthropic’s frontier reasoning model, designed for multi-step tasks that require deep planning and accuracy. It’s the model Cognition recommends for code review, architecture design, and debugging hard bugs. On FrontierCode 1.1 Main, Fable 5.1 scores roughly 62–63% (Cognition hasn’t published an exact figure, but previous announcements pegged it “within 2 points of Fable 5.0’s 63.2%”).
Sonnet 5.5, when running in extra-high reasoning mode, now scores 64.4% — slightly above Fable 5.1. The trade-off is speed: extra-high reasoning mode takes 2–3x longer to complete a task than standard reasoning (per Anthropic’s benchmarks), so it’s not the default setting. But for tasks where quality matters more than latency — like reviewing a critical pull request or debugging a production incident — Sonnet 5.5 at extra-high reasoning is now the best option in Desktop and CLI.
This matters for local-first workflows. Engineers who run Devin Desktop locally (rather than sending code to Cognition’s cloud) now have access to frontier-class reasoning without needing to route tasks through Devin Cloud. That’s a win for teams with compliance requirements (code can’t leave the local machine) or latency concerns (cloud round-trips add 500ms–2s per model call).
The pricing and quota implications
Sonnet 5.5 is priced at $3 per million input tokens and $15 per million output tokens via Anthropic’s API, per the company’s August 28 pricing update. That’s the same as Sonnet 5.0, but the extra-high reasoning mode effectively doubles token usage (since the model runs more internal reasoning steps before generating output). For Desktop and CLI users paying directly via Anthropic API keys, that means a typical coding task costs $0.10–$0.30 in Sonnet 5.5 standard mode, or $0.20–$0.60 in extra-high reasoning mode.
For ChatGPT Plus and Pro users who signed in to Devin via the September 29 ChatGPT quota integration, Sonnet 5.5 is not included — that integration only covers OpenAI models (GPT-4o, o1-preview, etc.). To use Sonnet 5.5 in Desktop or CLI, you need either a Cognition subscription (which bundles model costs) or a direct Anthropic API key.
Method: FrontierCode scores (64.4% for Sonnet 5.5, 56.2% for Sonnet 5.0) are from Cognition's September 28 announcement. The 500K token context window and extra-high reasoning mode details are from Anthropic's August 28, 2026 Sonnet 5.5 release notes. Fable 5.1's score range (62–63%) is inferred from Cognition's July 2026 announcement that Fable 5.1 performs "within 2 points of Fable 5.0's 63.2%." Sonnet 5.5 API pricing ($3 input / $15 output per million tokens) is from Anthropic's public pricing page as of August 2026. Task cost estimates ($0.10–$0.60) are based on typical FrontierCode task token counts (10K–20K tokens total) reported in Cognition's benchmark methodology docs.
References
- Sonnet 5.5 now in Devin Desktop and CLI — Cognition, September 28, 2026.
Editorial from Devin Central — news coverage by a Devin Ambassador, not Cognition. Devin is a trademark of Cognition; Claude is a trademark of Anthropic.