Two Claude deadlines hit today. Most teams aren’t ready.
At 9 AM Pacific, Anthropic is pulling the plug on two original Claude 4 model IDs with no grace period. Simultaneously, every programmatic Claude usage (Agent SDK, claude -p, GitHub Actions, third-party ACP clients) is being separated from the interactive subscription pool into a metered credit at full API rates.
If you’re building on Claude, you need to know about both. And you need to move today.
What dies today
Two model IDs are being retired, effective immediately:
claude-sonnet-4-20250514→ replaced byclaude-sonnet-4-6claude-opus-4-20250514→ replaced byclaude-opus-4-8
These were deprecated on April 14, which started Anthropic’s standard 60-day retirement clock. That clock runs out today. Calls to the deprecated model IDs will return errors immediately. There is no gradual wind-down.
This applies to Anthropic-operated platforms (the Claude API, Claude Platform on AWS, Microsoft Foundry). Partner platforms like Bedrock and Vertex AI set their own schedules, so if you’re running through those, check your provider’s docs.
There’s another one coming: claude-opus-4-1-20250805 retires August 5 and maps to the same Opus 4.8 target.
The Opus 4.8 trap nobody’s talking about
Here’s the detail that’ll trip up the most people.
Opus 4.8 (and everything back to Opus 4.7) dropped support for manually setting temperature, top_p, and top_k. If your code sends these parameters with non-default values, you’ll get a 400 error. Not a deprecation warning. Not a silent default. A hard error.
The fix: remove those parameters entirely from your API calls and use prompting best practices instead. If you’re migrating from Opus 4 to 4.8 today, this is the breaking change most likely to blow up your CI pipeline.
The Agent SDK billing split
This is the bigger story. Anthropic is splitting programmatic Claude usage out of the subscription pool and into a separate monthly credit.
Here’s what the new credit table looks like:
| Plan | Subscription | Agent SDK Credit | ~Tasks/Month |
|---|---|---|---|
| Pro | $20/mo | $20 | ~50 |
| Max 5× | $100/mo | $100 | ~250 |
| Max 20× | $200/mo | $200 | ~500 |
| Team Standard | $20/seat | $20/user | ~50/user |
| Team Premium | $100/seat | $100/user | ~250/user |
| Enterprise Standard | seat-based | $0 | 0 |
| Enterprise Premium | $200/seat | $200/user | ~500/user |
A few things jump out.
The task estimates assume Sonnet 4.6 pricing ($3/$15 per million input/output tokens) with a 10K input, 3K output workload. Real-world tasks like code review of a 1K-line PR, multi-file refactors, and eval suites burn through that credit faster than the estimates suggest.
And then there’s the Enterprise Standard surprise. If your company bought standard Enterprise seats, you get $0 in Agent SDK credit. Zero. You’d need Premium seats ($200/user) or to add usage-based seats ($20) to get any programmatic access at all. This is the most underreported detail in the whole change.
Interactive TUI, Cowork, and claude.ai are unaffected. The split only hits programmatic usage: Agent SDK, claude -p, GitHub Actions, and third-party ACP clients like Zed and JetBrains.
Zed put it plainly: “Claude subscriptions previously subsidized agent usage at roughly 15-30× compared to API pricing, and the new credits are billed at full API rates once depleted.” That subsidy is gone.
The credits are per-user (not pooled across teams), refresh monthly with your billing cycle, and do not roll over. You also need to opt in once per account. If your credit runs out and usage credits aren’t enabled, your requests just fail.
What to do today
Here’s the migration checklist. Run through it before your next CI pipeline fires.
- Export your current usage. Grep your config files, environment variables, and code for
claude-sonnet-4-20250514andclaude-opus-4-20250514. These are the model IDs that are about to break.
- Swap to the new IDs. Replace
claude-sonnet-4-20250514withclaude-sonnet-4-6. Replaceclaude-opus-4-20250514withclaude-opus-4-8.
- Remove sampling parameters on Opus. If you’re using Opus 4.8 and your code sets
temperature,top_p, ortop_k, delete those parameters. This is the one that catches people off guard.
- Check your plan’s Agent SDK credit. Figure out what your team gets under the new structure. If you’re on Pro, that’s $20 in credits. If you’re on Enterprise Standard, that’s $0.
- Test before your next CI run. Don’t find out your pipeline is broken at 9 AM because you didn’t swap the model ID in your config.
- Turn on prompt caching. This is the single highest-leverage cost optimization for Agent SDK users. Cache hits reduce input costs to 0.1× of the standard rate. If you’re not caching, you’re burning money.
The pricing math
If you’re wondering what $20 in credits actually gets you, here’s the breakdown from Digital Applied’s worked examples:
A code review of a 1K-line PR (30K input, 9K output tokens on Sonnet 4.6) costs about $0.225. That gives you roughly 89 code reviews from a Pro credit.
A multi-file refactor (120K in, 18K out) costs about $0.63. About 32 refactors per credit.
A full eval suite run (250K in, 50K out) costs about $1.50. About 13 runs per credit.
For context, a $20 Pro subscription running heavy Agent SDK loops used to access roughly $300-600 of API-equivalent compute. That 15-30× subsidy is what made the subscription feel like a steal for developers. Now you’re paying full API rates from a monthly pot that doesn’t grow with your usage.
Opus 4.8: the forced upgrade target
Opus 4.8 is genuinely impressive: 1M token context (up from 200k), 128k max output, adaptive thinking that allocates reasoning depth per turn, and Dynamic Workflows that can run hundreds of parallel subagents in a single session (research preview for Enterprise/Team/Max).
It scored 91.4% on Terminal-Bench 2.1 and is 4× less likely than its predecessor to let flawed code slip through. Databricks reports it’s 61% cheaper per multimodal token than Opus 4.7. I’ll do a proper deep dive when I’ve had real time with it.
The bigger picture
This isn’t just an Anthropic story. Ina Fried at Axios nailed it: “The fight shows that ‘all-you-can-eat’ AI subscriptions may not survive the agent era.”
Anthropic is the first major lab to hit this wall because it was the top choice for coding agents. The economics don’t work when your most valuable customers burn the most compute per dollar.
OpenAI is already exploiting the backlash, offering 2 months of free Codex to new business customers. Anthropic countered by bumping weekly usage caps by 50% for all paid tiers through July 13.
If you’ve been following the Anthropic export control saga, the Fable 5 and Mythos 5 forced suspension last week is part of the same story.
Andy Jassy’s role in triggering those controls and Dario Amodei’s policy response connect to the same tension: export controls, pricing pressure, and model lifecycle all at once.
I covered the Fable 5 launch and its usage-credit billing when it went public on June 9. The AI subscription landscape already had a callout for today’s billing split.
Bottom line
Today’s changes are real and immediate. Swap your model IDs, check your Agent SDK credit allocation, remove the temperature/top_p/top_k parameters if you’re on Opus, and turn on prompt caching. The all-you-can-eat era for AI subscriptions isn’t over, but the definition of “all you can eat” just got a lot narrower.
