Skip to main content

Claude 4.5 Sonnet Launches: 1M Context Goes GA and the Agent Era Accelerates

On September 29, 2025, Anthropic released Claude Sonnet 4.5, billing it as "the best model in the world for agents, coding, and computer use." In a single coordinated launch, Anthropic delivered not just a new model but a complete Agent toolchain — Claude Code upgrades, a native VS Code extension, the Claude Agent SDK, and Context Management APIs — and formally moved its 1M-token context window to general availability (GA) for all paid customers.

1. Positioning: The Trifecta of Agents, Coding, and Computer Use​

Anthropic's release notes stake three confident claims:

  1. Best coding model — top of the SWE-bench Verified leaderboard at launch;
  2. Best model for building complex agents — stable, multi-step autonomous execution;
  3. Best computer-use model — 61.4% on OSWorld, up from Sonnet 4's 42.2% just four months earlier.

The most striking internal metric is 30-hour sustained focus: in Anthropic's own tests, Sonnet 4.5 can drive a complex, multi-step task for more than 30 hours without human intervention. That single number demolishes the long-standing assumption that "AI agents are only good for short tasks."

2. Million-Token Context: From Beta to GA​

The 1M-token context window had been available in beta on Sonnet 4. With Sonnet 4.5, Anthropic graduates it to general availability (GA).

For enterprise users, this unlocks three production-grade patterns:

  • Drop an entire technical manual, full code repository, or lengthy legal filing into a single prompt;
  • Maintain precise recall of early-context details across long, multi-turn sessions;
  • Pack a whole project — code, historical issues, design docs — into a single inference call.

Million-token context changes how models consume information: context is no longer a scarce resource but a "working memory." That shift directly enables the new Context Editing and Memory Tool features discussed below.

3. Pricing and Availability​

Sonnet 4.5 keeps the exact API price of Sonnet 4:

  • Input: $3 per million tokens
  • Output: $15 per million tokens
  • Prompt caching: up to 90% cost savings
  • Batch processing: 50% cost savings
  • US-only inference: 1.1x pricing

This "big performance leap, no price hike" strategy is Anthropic's deliberate move to capture enterprise share. Sonnet 4.5 launched simultaneously on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry under the claude-sonnet-4-5 model name.

4. Companion Product Upgrades: The Agent Toolchain​

Sonnet 4.5 was not released in isolation — it ships with a full Agent operating stack.

4.1 Claude Code Upgrades​

  • Checkpoints — automatically snapshot code state before each change; revert with Esc Esc or /rewind;
  • Subagents — let a primary agent spin up specialized sub-agents in parallel (e.g., backend API while frontend is being built);
  • Hooks — trigger actions at specific points, such as running tests after code changes or linting before commits;
  • Background tasks — keep long-running processes (dev servers, watchers) alive without blocking the main agent loop;
  • Native VS Code extension — a dedicated sidebar panel with real-time inline diffs and editing suggestions.

4.2 Claude Agent SDK​

Anthropic packaged the core tools, context management systems, and permissions frameworks that power Claude Code itself into an SDK for third-party developers. Any team can now build custom agentic experiences on the same architecture Anthropic uses internally — a clear tailwind for startups in financial compliance, cybersecurity, and code debugging.

4.3 Claude for Chrome​

The browser extension is now available to Max subscribers. Combined with Sonnet 4.5's Computer Use improvements, agents can drive real browser workflows: competitive analysis, procurement, customer onboarding.

4.4 Context Editing and Memory Tool​

Two new context-management primitives let agents manage their own working memory — automatically pruning redundant information and persisting long-lived facts to external memory — so they can stay coherent across much longer tasks.

5. Benchmarks and Alignment​

  • SWE-bench Verified: state-of-the-art at launch;
  • OSWorld: 61.4% (vs. Sonnet 4's 42.2% four months prior);
  • Reasoning and math: significant gains across the eval suite;
  • Alignment: described by Anthropic as "the most aligned frontier model we've ever released," with measurable improvements across multiple alignment dimensions;
  • Expert blind review: domain experts in finance, law, medicine, and STEM rated Sonnet 4.5 substantially higher than earlier models — including Opus 4.1.

Customer quotes reinforce the picture: GitHub Copilot reported major gains in multi-step reasoning and code comprehension; Cursor validated improvements on longer-horizon tasks; Thomson Reuters' CoCounsel called it SOTA on the most complex litigation work; Block said Sonnet 4.5 cut the average vulnerability intake time for its Hai security agents by 44% and improved accuracy by 25%.

6. What This Means for Enterprises​

  1. GA long-context simplifies the engineering stack — many RAG-heavy or chunking-heavy pipelines can now be replaced by native long-context calls.
  2. A standardized Agent toolchain means enterprises no longer need to reinvent the loop, state management, and permission-isolation primitives; they can reuse Anthropic's already-validated Agent SDK.
  3. Zero-cost performance upgrade lets existing Sonnet 4 customers migrate with no contract change while gaining stronger coding, longer Agent tasks, and broader tool use.

7. Limits and Outlook​

Two caveats are worth flagging: 30 hours of sustained focus is not 30 hours of zero error — hallucination and goal drift remain real challenges in long-horizon agentic work; and Sonnet 4.5 did not deliver a notable breakthrough in multimodal generation (image or video).

Overall, Sonnet 4.5 is more than a model bump — it is the moment Anthropic turned the "Agent operating system" from concept into product-grade API. The signal to the industry is unambiguous: in the second half of the model race, Agent toolchains and context engineering will be just as decisive as raw model capability.

References:

  1. Anthropic Official: Introducing Claude Sonnet 4.5 (anthropic.com)
  2. Anthropic Official: Enabling Claude Code to work more autonomously (anthropic.com)
  3. Anthropic Models Overview: Claude Sonnet product line (anthropic.com/claude/sonnet)
  4. Anthropic Engineering: Building agents with the Claude Agent SDK