Who this is for
Founders and product leads who ship an in-product assistant, coding agent, or grounded chat where one model string is the user-facing default. If you only call models in offline batch jobs, this is not your week.
You are in the buyer if August price posts changed a finance or eng conversation, and someone asked whether to switch the default before the quarter closes.
What the published rates say (fetched 28–29 August 2026)
| Model | Input / output (per 1M tokens) | Status on the vendor page |
|---|---|---|
| OpenAI GPT-5.6 Sol | $4 / $20 | Promotional pricing available at least through 21 November 2026 (OpenAI developer pricing page) |
| OpenAI GPT-5.6 Terra | $2 / $12 | Listed on the same OpenAI pricing table |
| OpenAI GPT-5.6 Luna | $0.20 / $1.20 | Listed on the same OpenAI pricing table |
| Anthropic Claude Sonnet 5 | $2 / $10 | Published standard API rate on Anthropic’s pricing page |
| Anthropic Claude Opus 5 | $5 / $25 | Published standard API rate |
| Anthropic Claude Haiku 4.5 | $1 / $5 | Published standard API rate |
Reuters reported on 21 August 2026 that OpenAI cut GPT-5.6 Sol developer pricing by more than 20 percent for about three months, and that Plus / Pro / Business subscription prices were unchanged. Use the OpenAI pricing page as the contract text for API work; use Reuters only as confirmation that the cut was announced.
The trap is the cliff, not the screenshot
A default model change is a product change. It moves quality, latency, tool-calling behaviour, refusal style, and cost at the same time. A pricing screenshot moves only the cost cell, and only until the promo ends.
OpenAI’s own line is a floor on the discount (“at least through November 21, 2026”), not a published successor rate for 22 November. If your FY plan or customer pricing assumes $4 / $20 forever, you are budgeting a rate the vendor has not committed past that date.
Anthropic’s Sonnet 5 page currently shows $2 / $10 as the standard rate. That is the opposite problem: the rate looks permanent today, so the decision is quality and latency on your traffic, not a calendar cliff.
When switching the default is still the money move
Switch only if all three are true in writing:
- A fixed eval set (same prompts, same tools, same graders) beats the current default on the metrics you already ship against.
- Unit economics still work after you replace promotional rates with a conservative post-cliff assumption you wrote down.
- You can roll back the default string in under an hour without a store review or a contract change.
If any one fails, keep the current default. Route a shadow percentage to the candidate. Revisit after you have a week of production traces, not after a pricing blog.
This week
- Export last-30-day token mix by model, and the share that hits the product default path.
- Rebuild cost with published rates. Put Sol at $4 / $20 and a second column that assumes the promo ends after 21 November with no successor published.
- Freeze an eval set of at least 50 real prompts from production or support. Score current default vs one candidate. Do not cherry-pick demos.
- Do not change the user-facing default string this week unless the eval and the cliff-aware cost both clear your written bar.
- If finance asked because of the Reuters note, send them the OpenAI promotional sentence and the Anthropic Sonnet 5 standard rates, not a Twitter collage.
The public page is the call. The paid brief runs the same call against your token mix, eval scores, and whether a mid-tier (Terra / Sonnet / Haiku) already covers most traffic.
What I could not verify
I did not verify Google Gemini Flash rates from Google’s own pricing HTML in this pass (third-party trackers disagree on intro windows). I did not pull latency or quality numbers for Sol vs Opus 5 vs Sonnet 5 on a public leaderboard as a substitute for your eval. I did not confirm what OpenAI will charge for Sol on 22 November 2026; the page only guarantees the promo through at least 21 November. Subscription (ChatGPT Plus / Pro / Business) prices are out of scope for an API product default.
Sources
- OpenAI developer pricing page (platform.openai.com/docs/pricing), including the GPT-5.6 Sol promotional note through at least 21 November 2026 and the listed Sol / Terra / Luna rates. Fetched 29 August 2026.
- Anthropic, “Plans & Pricing” API section. anthropic.com/pricing#api — Sonnet 5 $2 / $10, Opus 5 $5 / $25, Haiku 4.5 $1 / $5. Fetched 29 August 2026.
- Reuters, “OpenAI cuts developer pricing for frontier GPT-5.6 Sol model by more than 20%,” 21 August 2026. reuters.com/.../openai-cuts-developer-pricing-frontier-gpt-56-sol-model-by-more-than-20-2026-08-21
Also on this site: Should I leave Apple IAP in the EU before 1 October 2026?, If I only call OpenAI or Claude, do I have GPAI duties?, and the Money Brief sales page.