Verdict: Anthropic has officially released Claude Sonnet 5.5 today, September 28, 2026. The new mid-tier frontier model delivers an unprecedented 70.6 percent score on Terminal-Bench 4.0—outperforming both earlier flagship Opus 5.5 and rival frontier systems—while maintaining permanent $2 per million input and $10 per million output token pricing. Alongside the general release across the Claude API, Claude Code, and GitHub Copilot, Anthropic has activated an on-demand “Reset for Free” feature in account settings to let subscribers immediately replenish session allowances.
Official Launch: Anthropic Releases Claude Sonnet 5.5 Today
Anthropic officially launched Claude Sonnet 5.5 (API model identifier claude-sonnet-5-5) today, September 28, 2026. Arriving less than one week after the September 22 debut of the flagship Claude Opus 5.5, Sonnet 5.5 completes the rapid transition of Anthropic’s production tier to the 5.5 architecture. The model is now generally available across the Anthropic Console, public API endpoints, Claude Web and Desktop applications, Claude Code, GitHub Copilot, and enterprise cloud marketplaces including Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure.
The release confirms the intense developer speculation and terminal reports that circulated throughout the morning. Earlier observations of canary routing and dark testing within Claude Code CLI sessions were indeed the final staging gates for today’s widespread public rollout. Sonnet 5.5 positions itself as the primary high-throughput engine for software development, terminal task automation, continuous integration triage, and enterprise document generation.
The Benchmark Breakthrough: Sonnet 5.5 Tops Terminal-Bench at 70.6%
The defining technical milestone of Claude Sonnet 5.5 is its performance in autonomous coding and command-line execution. In Anthropic’s published evaluation suite, Sonnet 5.5 recorded a score of 70.6 percent on Terminal-Bench 4.0:
- Surpassing Flagship Opus 5.5: Remarkably, Sonnet 5.5’s 70.6 percent score exceeds the 66.4 percent mark posted by Claude Opus 5.5 just six days prior. This makes Sonnet 5.5 Anthropic’s highest-scoring model for terminal command synthesis and autonomous coding agents.
- Generational Leap Over Sonnet 5: Compared to its direct predecessor, Claude Sonnet 5 (which scored 10.3 percent on Terminal-Bench 4.0 upon its June 30 release), Sonnet 5.5 demonstrates a near seven-fold jump in command-line reasoning, error correction, and multi-file code editing.
- Outperforming Competitor Frontier Models: Sonnet 5.5 comfortably outpaces OpenAI’s GPT-6 Astra (57.9 percent) and GPT-6 Sol in terminal agent tasks, while running at a substantially lower per-token price point.
- Broad Occupational Reasoning (GDPval-AA): On the GDPval-AA benchmark, which evaluates professional problem-solving across specialized occupations, Sonnet 5.5 performed within two points of Opus 5.5, demonstrating that its reasoning depth extends well beyond raw syntax generation.
This leap in terminal competency reflects an industry-wide transition toward autonomous agentic tooling, a trend also seen in Google’s internal development cycles analyzed in our report on Gemini 4 Pro specifications and benchmark leaks.
Pricing and Economics: $2/$10 Pricing with 30% Net Task Savings
Despite the substantial gains in benchmark accuracy, Anthropic decided against increasing token rates. Claude Sonnet 5.5 preserves the exact permanent pricing established by Sonnet 5:
- Input Tokens: $2.00 per million tokens.
- Output Tokens: $10.00 per million tokens.
- Prompt Cache Reads: $0.20 per million tokens.
- Prompt Cache Writes: $2.50 per million tokens.
Critically, Anthropic reports that organizations will experience an effective cost reduction of up to 30 percent per completed task. Because Sonnet 5.5 resolves errors with fewer exploratory tool iterations and generates output code with greater conciseness, it consumes fewer overall input and output tokens to complete identical programming tickets. Combined with output token generation speeds exceeding 30 percent faster than the 5.0 baseline, software engineers receive immediate latency and budgetary relief.
Anthropic’s “Reset for Free” Launch Bonus: How to Claim It in Settings
To celebrate the release and assist subscribers who want to stress-test Sonnet 5.5 without running into message caps, Anthropic has activated an official “Reset for free” capability for eligible paid plans (Claude Pro, Max, and Team tiers).
Step-by-Step Claim Instructions
The reset is a manual trigger accessible through the web and desktop interfaces:
- Navigate to Claude on your web browser or open the Claude Desktop application.
- Click your profile icon in the lower-left corner and choose Settings.
- Select the Usage tab from the navigation sidebar.
- Scroll to the Resets module. Under active launch campaigns, you will see a button labeled “Reset for free” with an associated expiration timestamp.
- Click the button and confirm the modal prompt to immediately restore your session and rolling allowances to 100 percent.
Operational Rules and Terminal Synchronization
Keep these verified guidelines in mind when planning your usage:
- Full Account-Wide Synchronization: While the button must be pressed in the web browser or desktop app, the quota replenishment applies account-wide. If you are developing inside Claude Code or using an IDE extension and hit a rate limit, triggering the reset in your browser immediately restores your terminal coding capacity.
- Fixed Expiration Windows: Campaign resets carry a defined expiration date (typically 30 days from release). If left unclaimed, the bonus credit expires automatically.
- Irreversible Execution: Triggering the reset cannot be paused or undone. Use it strategically when beginning an extensive multi-file refactoring sprint or large test-suite generation.
Claude 5.5 Family Specifications & Market Comparison
The table below provides a side-by-side comparison of confirmed specifications across Anthropic’s updated model roster and leading market alternatives:
| Model | Availability Status | Terminal-Bench 4.0 | Pricing (Input / Output per 1M) | Context Window | Primary Specialization |
|---|---|---|---|---|---|
| Claude Sonnet 5.5 | Live Today (Sep 28, 2026) | 70.6% | $2.00 / $10.00 | 1,000,000 tokens | High-speed agentic coding, command-line automation, CI/CD triage |
| Claude Opus 5.5 | Live (Sep 22, 2026) | 66.4% | $4.00 / $20.00 | 1,000,000 tokens | Complex systems architecture, theoretical reasoning, deep research |
| Claude Sonnet 5 | Superseded (Jun 30, 2026) | 10.3% | $2.00 / $10.00 | 1,000,000 tokens | General developer workflows (now upgradeable to Sonnet 5.5) |
| GPT-6 Astra | Active Competitor | 57.9% | $5.00 / $22.50 | 1,000,000 tokens | Multimodal reasoning, cross-application enterprise workflows |
| Claude Fable 5.1 | Active (Specialized) | N/A (Theoretical) | $8.00 / $40.00 | 500,000 tokens | Long-horizon multi-hour agent operations, formal verification |
Cybersecurity Safeguards and Visual Execution (Beating Pokémon Red)
Anthropic equipped Sonnet 5.5 with two unique capability advancements previously restricted to proprietary research environments:
1. Frontier-Grade Cybersecurity Protection
Sonnet 5.5 is the first model in the Sonnet tier to deploy with Anthropic’s advanced cyber-defense classifiers. These safeguards actively detect and block attempts to generate exploit payloads, reverse-engineer proprietary security boundaries, or extract underlying reasoning traces. This protection provides enterprise teams with confidence when granting the model elevated shell permissions in automated testing environments.
2. Pure Visual Decision-Making: The Pokémon Red Milestone
In a notable technical demonstration of visual grounding, Anthropic verified that Sonnet 5.5 is the first Sonnet-class model capable of playing and completing Pokémon Red exclusively through raw screen captures. The model interprets pixel data, navigates menu hierarchies, tracks battle mechanics, and issues directional controls without access to internal game memory or text stream hooks. This visual fidelity translates directly into real-world utility for automated web application testing and GUI navigation.
Developer Implementation: How to Migrate Your Workflows Today
For engineering teams operating API integrations, migration to Claude Sonnet 5.5 requires minimal configuration adjustments:
- Update Model Identifiers: In your API payload headers or client configurations, replace
claude-sonnet-5-20260630withclaude-sonnet-5-5. In Python or TypeScript SDKs, this is a single string substitution. - Configure Claude Code: Run
claude updatein your terminal to ensure you are operating on CLI version 2.1.280 or newer. Verify your connected model by typing/statusinside an active session. - Adjust Thinking Effort Budgets: Because Sonnet 5.5 exhibits high baseline reasoning density, set your thinking parameter to
lowormediumfor standard code reviews and documentation generation. Reservehighormaxeffort strictly for full-repository refactors and architectural migration scripts. - Integrate with Developer Suites: Sonnet 5.5 is immediately selectable in GitHub Copilot’s multi-model developer suite, allowing individual contributors to toggle Sonnet 5.5 alongside OpenAI and Google models directly within their editor.
For developers architecting broader organizational automation stacks, explore how this fits with long-term infrastructure in our breakdown of Claude Fable 5 and Mythos 5 deployment strategies, and discover complementary tooling in our guide to the 10 best AI productivity tools for automating developer workflows.
Summary and Next Milestones
The official launch of Claude Sonnet 5.5 today cements Anthropic’s lead in agentic software engineering. By delivering a category-leading 70.6 percent Terminal-Bench score at an unchanged $2/$10 price point, Sonnet 5.5 establishes a new cost-to-performance standard for developers worldwide. With Opus 5.5 and Sonnet 5.5 now live, all eyes turn to the final member of the family—Claude Haiku 5.5—expected to complete the generational rollout in the coming weeks.