The short version: two dated events in the next four weeks change what you pay to run AI, and both land on a team of one. On August 26, OpenAI's Assistants API stops responding — ignore it and your agent goes dark. On September 1, Claude Sonnet 5 costs 50% more. Neither is optional; each has a clean fix; and one well-placed abstraction handles both. Here's the money math and the move.
| Event | Date | What it does to your bill | The fix |
|---|---|---|---|
| Assistants API sunset | Aug 26 | Nothing — until it breaks (hard) | Migrate to the Responses API (GA) |
| Sonnet 5 intro pricing ends | Aug 31 → Sep 1 | +50%, uniform | Route bulk to a cheap tier |
August 26 — the OpenAI Assistants API stops responding#
This one isn't a price change; it's a cutoff. On August 26, 2026, every call to /v1/assistants, /v1/threads, and /v1/threads/runs returns an error — no grace period, no read-only mode, no auto-forwarding (OpenAI). If your agent still speaks the Assistants dialect, its "cost" on the 26th is an outage.
The fix: move to the Responses API, which is already GA. It's a mental-model shift, not a rewrite — the old four-object dance (Assistant, Thread, Messages, Run) collapses into a single client.responses.create() call, and your vector stores carry over free (reuse the same vs_… IDs with file_search, no re-upload). The only real work is discovering which tool calls still assume a Thread, so grep for beta.threads today, not on the 25th. The endpoint-by-endpoint mapping with code is here.
September 1 — Sonnet 5 costs 50% more#
Claude Sonnet 5 launched June 30 at an introductory $2/M input, $10/M output. That rate ends August 31; from September 1 it's $3/$15 — a flat 50% rise applied uniformly to base, both cache-write tiers, cache reads, and the 50%-off Batch API rate (Anthropic pricing). No amount of caching or batching claws it back, because those discounts apply to a base rate that itself went up.
The concrete bite: a mid-size agent burning 500M input + 25M output tokens a month goes from $1,250 to $1,875 — an extra $625/month, $7,500 a year, for identical work.
The fix: re-price your backend before month-end. The two levers that actually beat the hike are sending fewer tokens or changing tier, so default new bulk volume — extraction, classification, background loops — to a cheap tier and keep Sonnet 5 for the reliability-critical paths that earn it. This week's Wire made the same case after DeepSeek V4 Flash out-benchmarked flagships; the head-to-head math is in DeepSeek V4 Flash vs Sonnet 5 before the price cliff, and the exact bill impact — including the Sonnet 4.6 tokenizer wrinkle — is here.
The one move that answers both#
Notice what both fixes have in common: they're easy if your provider is a setting and painful if it's hard-coded. So the highest-leverage afternoon this month is to put a swappable client in front of your LLM calls — a thin layer where the provider, model, and base URL are configuration, not scattered call sites.
Do that and the Assistants migration becomes a one-file change instead of a codebase hunt, and the Sonnet price flip becomes a routing decision you make in a config, not a scramble. Route bulk work to the cheap tier, keep the premium model on the paths that need it, and A/B the two on your own eval — task success, tool-call validity, cost per completed task — before September 1. If you're starting from scratch, how to build a cost-aware model router is the pattern.
That's the whole playbook: one migration that can't slip past the 26th, one price rise that lands on the 1st, and one abstraction that turns both from fire drills into config. Two of these deadlines have real money attached; the fix for both is the same afternoon. (For the fuller list of what else lands this quarter — the EU marking rule and the China data window — the founder's calendar tracks the dates.)
