The past three days brought a concentrated release from Anthropic's developer platform: five tools graduated from beta in a single morning, followed the next day by a major SDK version that carries real migration overhead. Alongside those changes, a pricing commitment simplifies budget planning for the rest of the year, and a governance milestone is quietly setting the protocol architecture that will underpin enterprise multi-agent work for the next cycle.

Anthropic graduates five agentic developer tools from beta in a single release

On 19 August, Anthropic removed beta-header requirements from five separate developer APIs simultaneously. That is an unusual concentration of graduation events, and each one matters on its own terms.

The computer use tool is now generally available as computer_toolset_20260801. Alongside graduation, it picks up batch actions (multiple interactions in a single turn), zoom enabled by default, and per-member configuration through the configs array. Teams migrating from the earlier computer_20251124 version need to update request shapes and tool-handling logic.

Launching simultaneously is a brand-new browser use tool (browser_toolset_20260801). Unlike computer use, which operates on a full desktop, browser use works inside a browser viewport that your application hosts. It reads the page's accessibility tree, forms, and tabs directly rather than relying on screenshots, which makes it substantially more robust for web-automation workflows. Element references, form input, tab management, download reporting, and optional file upload are available as first-class operations.

The three further GAs are platform-plumbing that builders have been holding in beta:

  • The Files API is now stable, with file expiration (expires_in_seconds at upload; expires_at on file objects), page-based pagination, and an ids[] filter when listing files.
  • Agent Skills and the /v1/skills endpoint are GA: the skills-2025-10-02 beta header is no longer required.
  • The Admin API for Claude Enterprise user management (members, invites, groups, custom roles) is GA for all Enterprise organisations.

For operators, the practical implication is direct: each of these was a reasonable reason to hold back a production build. All five reasons are now gone.

Python SDK v1.0 ships with breaking changes that require migration

On 20 August, Anthropic released v1.0 of the Python SDK. The headline infrastructure change is the HTTP layer moving from httpx to httpx2, a maintained API-compatible fork. Teams using tracing or mocking libraries that patch httpx will need to call httpx2.alias_httpx() at startup. Python 3.10 or later is required.

The deeper breaks are in the API surface itself:

  • The legacy Text Completions API is removed entirely.
  • temperature, top_p, and top_k are removed from Messages method parameters. The effort parameter is now the sole mechanism for steering model behaviour.
  • compaction_control is removed from the client-side tool runner.
  • On the async client, .with_raw_response results now require await response.parse().
  • AnthropicBedrock raises an error when no AWS region is configured, rather than defaulting to us-east-1.

Anthropic has published a migration guide with before-and-after examples for every change. Any team running automated pipelines via the Python SDK should test against v1.0 in a staging environment before promoting to production, particularly if sampling parameters are set explicitly anywhere in the codebase.

Claude Sonnet 5 pricing is now permanently $2/$10 per million tokens

Effective 10 August, Anthropic confirmed that the introductory price for Claude Sonnet 5 is the standard price. The increase to $3 per million input tokens and $15 per million output tokens, which had been announced as scheduled for 1 September 2026, will not occur.

The structural consequence: Claude Sonnet 5 is now 33 percent cheaper per input token and per output token than Claude Sonnet 4.6 ($3/$15), while offering higher capability across most benchmarks. Teams running Sonnet 4.6 as their production workhorse have a straightforward cost-reduction case for migrating, with no pricing headwind to plan around. Budget planning for Q3 and Q4 can treat the $2/$10 rate as settled.

Google's A2A protocol moves to the Agentic AI Foundation, alongside MCP

On 17 August, Google transferred the Agent2Agent (A2A) protocol to the Agentic AI Foundation (AAIF), the body that already hosts Anthropic's Model Context Protocol. The Foundation now counts more than 250 members, up from fewer than 40 at its December 2025 founding, with backing from Google, Microsoft, Amazon, Anthropic, OpenAI, Bloomberg, Shopify, and Block. Google reports 150 organisations routing real production tasks over A2A, not just running pilots.

The two protocols serve distinct roles that are worth keeping clear. MCP governs how an agent connects to tools and data sources. A2A governs how agents communicate with each other across platform and organisational boundaries. Having both under the same neutral foundation matters for enterprise buyers who need cross-vendor interoperability without a dependency on any single lab's continued stewardship of the standard.

For teams building multi-vendor agent architectures, the governance picture is now cleaner than it has been at any point in the past year. Two open protocols, one foundation, broad industry backing, and 150 organisations already running A2A in production: the window for hedging with proprietary alternatives is closing.

The through-line across all four items is maturation rather than novelty. Anthropic's five-item GA release, the v1.0 SDK, and the A2A governance shift all signal a transition from experimental infrastructure to production commitments. For operators, the question is no longer whether to build on agentic foundations; it is whether your engineering team's architecture and vendor contracts are keeping pace with how fast those foundations are now stabilising.