Starting June 15, 2026, Anthropic is ending the flat-rate subsidy that made heavy Claude agent use economical on subscription plans. Every programmatic call to Claude via agent SDK, headless claude -p mode, or GitHub Actions will draw from a hard monthly credit pool billed at standard API rates.
What is changing
Until now, a $100/month subscription gave you roughly 10x the token value at API rates. A $200/month subscription gave you 20x. That subsidy ends June 15.
From June 15, Claude charges API token usage at standard rate card prices whether you are on a subscription or not. Your plan tier only determines the size of your monthly credit pool, not the rate.
The exceptions: Claude.ai chat, Claude Cowork, and interactive Claude Code still receive subsidized usage as part of your subscription. Everything else, including all agent sessions running on your CRHQ satellite, moves to metered billing.
| Plan | Monthly price | SDK credits | At Sonnet rates | At Opus rates |
|---|---|---|---|---|
| Pro | $20 | $20 | ~4.4M tokens | ~2.7M tokens |
| Max 5x | $100 | $100 | ~22M tokens | ~13.3M tokens |
| Max 20x | $200 | $200 | ~44M tokens | ~26.6M tokens |
When credits run out, agent calls stop entirely (or continue at full API rates if you enable overflow billing). Required action: claim your SDK credit allocation before June 15 via the email Anthropic is sending to your account. See the official Anthropic announcement.
The most important thing to understand
The per-token API rate is identical at every Claude subscription tier. There is no discount for paying more. A $20 plan gives $20 in monthly credits at the same rate as a $200 plan gives $200. If you are not a heavy interactive Claude user, there is no reason to be above $20 per month.
What we recommend
Switch to OpenAI Codex (GPT-5.5)
This is the clearest like-for-like replacement. GPT-5.5 is directly comparable to Claude Opus 4.8 in intelligence. The Codex CLI authenticates through your ChatGPT Pro subscription using a rolling window allocation, not a hard monthly dollar cap.
Why GPT-5.5 via Codex:
- Comparable intelligence to Claude Opus 4.8 (see benchmarks below)
- Still heavily subsidized via ChatGPT Pro rolling window subscription
- More token-efficient: completes equivalent tasks using significantly fewer output tokens
- Multiple subscription tiers available: start lower, upgrade if you hit rate limits
- OpenAI has not announced plans to remove this subsidy
- Already integrated in your CRHQ satellite
Not using Claude Code interactively? Here is what to do before June 15:
- Open a ChatGPT Pro subscription at chat.openai.com/settings/subscription. Start with a lower tier and upgrade if you hit rolling rate limits.
- In your CRHQ satellite, go to Settings, then System and connect your ChatGPT subscription.
- Switch your Agents from Claude to OpenAI models and select GPT-5.5.
- Downgrade your Claude plan to Pro ($20/month) to keep the $20 SDK credit for remaining Claude tasks.
Heavy Claude Code user? Claim your credits first (via the Anthropic email), then try Codex side-by-side for agent work. For full details: Using Codex with your ChatGPT plan.
Cursor Composer 2.5: the most cost-effective option
Released May 18, 2026, Cursor Composer 2.5 is Cursor's own fine-tuned coding model (built on Kimi K2.5 with proprietary reinforcement learning). It is already integrated in your satellite.
- $0.50 input / $2.50 output per million tokens. That is 10x cheaper than Claude Opus 4.8.
- 79.8% on SWE-Bench Multilingual, near-parity with Claude Opus 4.7.
- Cursor Pro at $20/month includes a $20 credit pool. Cursor Ultra ($200/month) includes roughly $400 in credits.
How to set up Cursor Composer 2.5:
- Create a Cursor account at cursor.com and subscribe to Cursor Pro at cursor.com/pricing.
- In your CRHQ satellite, go to Settings, then System.
- Connect your Cursor account or paste your Cursor API key from your cursor.com account settings.
- Switch provider to Cursor and select Composer 2.5 as your model.
More on Composer 2.5: Cursor blog: Introducing Composer 2.5.
Benchmark comparison
The charts below compare all four models on SWE-Bench Verified (real GitHub issue resolution) and Terminal-Bench 2.0 (end-to-end CLI and shell agent tasks in Docker containers, directly relevant to how CRHQ satellite runs agents).
Side-by-side comparison
| Model | Monthly cost | SWE-Bench | API cost in/out per MTok | Recommended for |
|---|---|---|---|---|
| Claude Opus 4.8 | $200 Max (incl. $200 credits) | 88.6% | $5 / $25 | Tasks requiring deepest reasoning |
| Claude Sonnet 4.6 | $200 Max (incl. $200 credits) | 79.6% | $3 / $15 | Best Claude value post-June 15 |
| GPT-5.5 via Codex | $200 ChatGPT Pro (rolling window) | 88.7% | $5 / $30 API | Primary recommendation for most users |
| Cursor Composer 2.5 | $20 Cursor Pro ($20 credits) | 79.8% | $0.50 / $2.50 | Best cost-performance ratio |
ChatGPT Pro subscription uses rolling window, not per-token billing. Cursor Composer 2.5 scored on SWE-Bench Multilingual (not Verified).
A note on Claude Haiku
Claude Haiku 4.5 at $0.80 / $4.00 per million tokens is the cheapest Claude option for lightweight tasks. Not recommended for complex agentic work, but useful for quick lookups and high-volume simple automations.
Further reading
- Official Anthropic announcement: Agent SDK billing change on June 15
- Using Codex with your ChatGPT plan
- OpenAI Codex
- Cursor pricing and plans
- Cursor blog: Introducing Composer 2.5
- Terminal-Bench 2.0 leaderboard
- BenchLM.ai leaderboard
- llm-stats.com Terminal-Bench dataset
- Anthropic billing change explained
We will continue monitoring performance and cost across all providers and update this article as the situation evolves after June 15. Questions? Reach us at hello@crhq.ai.