AI API pricing cliffs in September 2026 — Sonnet stay and Copilot promo pools

One calendar cliff still hits AI buyers around 1 September 2026. The other — Claude Sonnet 5’s planned list step — will not. This note covers the canceled Sonnet cliff and the Copilot Business/Enterprise promo AI-credit pool drop, with a short checklist so engineering and finance share one forecast.

It does not replace vendor quotes, committed-use deals, or BYOK harness TCO models. Use it to stop “August run-rate forever” thinking where a cliff still exists, and to stop budgeting a Sonnet increase that Anthropic already canceled.

Claude Sonnet 5 — $2/$10 is now standard (cliff canceled)

Sonnet 5 launched with intro list pricing of $2 / $10 per million input/output tokens through 31 August 2026, with a scheduled step to $3 / $15 on 1 September. Anthropic later made the intro rate the standard price. The previously scheduled increase will not occur.

What that means for budgets:

• Direct Anthropic API spend can keep using $2 / $10 list for Sonnet 5 after 1 September (verify on Anthropic’s pricing docs before locking a forecast).
• BYOK harnesses that route to Sonnet 5 do not need a September list re-forecast for this model alone.
• Decks that still warn about a +50% Sonnet step on 1 September are stale — update them.

What still matters: Sonnet remains expensive per DeepSWE task versus Opus on some workloads even when per-token list looks mid-tier. Catalog context: Claude Sonnet 5. Broader model routing: GPT-5.6 Terra / Sol / Luna and the AI release watchlist.

GitHub Copilot org promo pools — after 1 September 2026

Business and Enterprise promo AI-credit pools (3,000 / 7,000) move to standard 1,900 / 3,900 after 1 September 2026. Individual Pro+ / Max stories are separate — do not mash consumer Copilot marketing into org pool maths.

What to do: open the org admin usage view and chart August burn against the post-promo pool. If you were coasting on promo headroom for agentic features, either cut agent defaults, buy additional credits deliberately, or move heavy agent work to a harness with clearer BYOK metering. See GitHub Copilot harness and setup.

Promo cliffs feel small until ten teams enable coding agents the same week. Treat 1 September as a soft launch gate for new Copilot agent rollouts unless the post-promo pool still clears your forecast with margin.

What else sits near the same calendar

• Claude Code’s temporary weekly usage boost ends 19 August 2026 — size seats without it (Auto mode note).
• ChatGPT Free/Plus Chat refreshed on 6 August — do not confuse Chat seats with API cliffs (Free vs Plus).
• Model EOLs and provider watch items remain on the sources / watchlist page — pricing cliffs and shutdowns both break naive forecasts.

A one-week finance + eng ritual

1. Export August token and Copilot credit burn by team.
2. Apply post-promo Copilot pools; leave Sonnet 5 at $2 / $10 unless your contract says otherwise.
3. Mark workloads that can move to cheaper model IDs without quality loss.
4. Freeze net-new Copilot agent rollouts that depend on promo surplus until owners accept the new envelope.
5. Update any public or sales “AI cost” claims that still cite a September Sonnet list step.

Common mistakes

• Still quoting a Sonnet $3/$15 September cliff after Anthropic canceled it.
• Assuming Copilot individual plans describe Business/Enterprise pools.
• Treating BYOK harness spend as “included in the IDE seat.”
• Ignoring FX when converting USD list cards for UK budgets.
• Planning Claude Code capacity on a boost that expires mid-August.

Bottom line

Budget Copilot org pools at 1,900 / 3,900 after the promo. Keep Sonnet 5 at $2 / $10 list unless your quote differs. Re-forecast this week where promo maths still matter, rematch volume work to cheaper IDs where validators already pass, and keep Chat, API, and harness meters separate. For coding seat decisions, use the AI coding assistant finder instead of buying more of the wrong surface.