Anthropic retires Claude Sonnet 4 and Opus 4 on the API

Anthropic

API · Model deprecations

Claude Sonnet 4 and Opus 4 reach end of life on the API today

As of 9am Pacific this morning, claude-sonnet-4-20250514 and claude-opus-4-20250514 are gone from the Anthropic API. There is no grace period and no gradual wind-down: an unmigrated integration simply returns an error from now on. If your automated workflows, scripts or product calls still reference either model string, today is the day those calls fail.

The fix is straightforward: repoint Sonnet 4 traffic to claude-sonnet-4-5 and Opus 4 traffic to claude-opus-4-5 (or the latest versions available on your plan), both of which are more capable and in several cases cheaper per token. Consumer Claude.ai accounts and managed Claude Code environments are unaffected because Anthropic selects the model on your behalf. If you build directly on the API, check your model strings today rather than discover the failure in production.

Sources

Anthropic Docs, 15 Jun Enterprise DNA, 15 Jun

Billing · Agent SDK

Agent SDK and headless usage move to a separate credit pool today

From today, the Claude Agent SDK and claude -p headless usage no longer draw from the Pro, Max, Team or Enterprise subscription pools. Both now bill against a separate monthly dollar credit at standard API rates, and that credit does not roll over at the month boundary. Unused allowance is lost. If you run agents or scripted headless jobs on the back of a subscription, both your effective limits and your costs change from this morning.

The practical consequence is that programmatic and agentic workloads are now metered distinctly from interactive chat. Jobs will stop when the credit pool runs dry unless you have overflow billing enabled, so checking your billing setup today is the immediate action item if you want to avoid surprise throttling or charges later in the month. Third-party agent apps built on the SDK (including OpenClaw, Conductor, Zed and Jean) are also affected. Interactive use on web, desktop, mobile, terminal Claude Code and Claude Cowork remains on the standard subscription.

Sources

Enterprise DNA, 15 Jun

Quiet today

OpenAI Nothing significant in the last 24 hours. Talk of a GPT-5.6 launch is tied to prediction markets and unofficial signals, with no official announcement, so it falls below the threshold.

Google DeepMind Nothing shipped in the last 24 hours. Gemini 3.5 Pro, shown at Google I/O with a 2-million-token context window and Deep Think mode, remains in limited Vertex AI enterprise preview with no public general availability date.

xAI Nothing significant in the last 24 hours. Grok Speech-to-Text and Text-to-Speech APIs and the Grok Build plugin marketplace are recent but landed earlier in the cycle, outside this window.

Meta Nothing significant in the last 24 hours. A five-year, $27 billion Nebius compute agreement is a recent infrastructure story, but it is a procurement deal rather than a product and predates the window.

Microsoft Nothing significant in the last 24 hours. Satya Nadella's widely shared commentary on owning your AI learning loops is editorial rather than a product launch.

Nvidia Nothing significant in the last 24 hours. The RTX Spark personal-AI chip and the NAVER gigawatt data centre partnership both predate the window.

Mistral · DeepSeek · Qwen (Alibaba) Nothing qualifying from any of these in the last 24 hours. Note for budget holders: Alibaba Cloud's recent price increases of up to 34% on compute and GPU instances are ongoing but not a same-day product change.

IBM · Snowflake Nothing significant in the last 24 hours.

Cohere · Perplexity · Amazon Bedrock · Hugging Face · Runway · ElevenLabs · Stability AI · Scale AI Nothing qualifying met the product-first, 24-hour threshold. ElevenLabs shipped minor API housekeeping this month (a revoke API key endpoint and higher speech-to-text upload limits) but none rose to a same-day launch.

Today's themes

The product front was genuinely quiet once speculation and regulatory commentary are filtered out. The one area where things actually land in your lap today is Anthropic's API: the Claude Sonnet 4 and Opus 4 retirements and the Agent SDK billing split both take effect this morning, making today a migration day for anyone building directly on Anthropic. The broader pattern these two changes reinforce is that the economics and operational contracts of using AI are tightening: model versions have a defined shelf life of roughly 90 days after a successor lands, and the era of flat-rate programmatic access to frontier models is ending. Separately, two anticipated releases, Gemini 3.5 Pro and a possible GPT-5.6, remain just out of reach, a reminder that much of the current moment is waiting rather than shipping.