AI API pricing cliffs — September 2026: the dated price steps to diarise

Price and product status checked on 1 October 2026. Dates below identify a specific scheduled change; current alternatives and model launch status are updated in the relevant sections.

A pricing cliff is a dated price step: a promotion ending, an intro rate expiring, a pool shrinking. None of them is a surprise if it is in the diary. Here are the ones that matter for budgets being set this autumn.

The cliffs to diarise

GPT-5.6 Sol is guaranteed promotional pricing at least through 2026-11-21. OpenAI lists the current API rate as $4/$20 and describes that price as promotional, but its model page does not announce an end date or post-promotion rate. Do not budget an automatic step to $5/$30 after 21 November without a new vendor notice. GPT-6 Sol launched on 22 September at $2/$10, half the current GPT-5.6 Sol promotional rate. If you migrate, re-test your workload and exact API model ID first. Migrating from GPT-5.6 to GPT-6 Sol and Luna has the checklist.

GPT-6.1 Sol is the newer OpenAI Sol model. OpenAI lists the same $2/$10 standard token rates as GPT-6 Sol, but cached input is $0.10/MTok instead of $0.20. The 272K long-prompt price step still applies. Compare the exact model ID and cache share before estimating savings. GPT-5.6 Sol’s documented price is promotional and available at least through 21 November; no later rate is announced.

Gemini 3.7 and 3.8 Flash API rates step on 2027-01-01. Google’s current Gemini API pricing page lists $0.75/$3.75 through 2026-12-31, then $1.50/$7.50. Model availability and pricing are separate: Gemini 3.7 remains listed as a previous-generation stable model with no shutdown date announced in the deprecations page checked 1 October.

DeepSeek has a daily cliff. DeepSeek bills peak and off-peak rates. V4.1 Flash is $0.30/$1.20 at peak and half that off-peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except Chinese public holidays; those holidays are off-peak for the full day. A batch job that drifts into peak hours doubles its cost. DeepSeek also retired V4 Flash on 10 September. The old deepseek-v4-flash name is still accepted, but routes to V4.1 Flash and bills at its rate.

Muse Spark 1.3 has a data-for-price choice. Meta’s current pricing docs list Standard at $1.25/$4.25 per MTok and say prompts and completions are not used to train Meta models. Contributor is $0.10/$0.20 and gives Meta permission to use prompts and completions to train future models. Check the tier-specific rates and data terms before choosing.

Every row in the model table shows the date it was last checked, so a stale forecast is visible at a glance.

Claude: Sonnet 5.5 is released at the same token rates

Claude Sonnet 5.5 launched on 28 September at $2/$10 per million input/output tokens, the same token rates as Sonnet 5. Anthropic says it is 30%+ faster and can cost up to 30% less per task for most work; those are vendor claims, not a guarantee for every workload. The API listing also shows $0.20 cache reads. Sonnet 5 remains in Anthropic's prior-model pricing table.

The remaining watch item is Claude Haiku 5.5, which Anthropic says will follow in the coming weeks; no release date or price is published. Do not forecast its price from another model's discount. Claude Opus 5.5 ($4/$20) remains the current Opus model, replacing Opus 5 ($5/$25) on 22 September.

Catalog context: Claude Sonnet 5.5 is the current Sonnet release; Claude Sonnet 5 remains a previous model row.

Google: a Gemini 4 watch note

Gemini 4 Argon is a restricted preview, not a generally available API alternative. Google has not published a purchasable API price for it. Treat any release timing or general-availability forecast as unconfirmed, and check Google's model catalog before making an annual Gemini 3.x commitment.

GitHub Copilot: credit pools and transition cases

GitHub announced that Copilot plans would transition to GitHub AI Credits on 1 June 2026. Current docs show individual monthly allowances and pooled Business/Enterprise credits. Some existing Pro/Pro+ annual subscribers may stay on legacy premium-request billing until their annual plan ends. For organizations, current docs list standard allowances of 1,900 credits per Business seat and 3,900 per Enterprise seat; confirm the active allowance and billing cycle in the admin usage view. Copilot Business is $19 per granted seat per month with 1,900 credits per user per month. Copilot Enterprise is $39 per seat with 3,900 credits. Code completions stay unlimited on paid plans and do not draw credits.

Open the org admin usage view and chart August burn against the current 1,900 / 3,900 standard allowance; confirm any account-specific transition or promotional balance there. Do not mix individual-plan allowances (Pro, Pro+, Max) into org pool maths. See the GitHub Copilot harness page for setup, and AI coding assistant pricing for how to forecast credit burn.

A one-week finance + eng ritual

1. Export August and September token and Copilot credit burn by team.
2. Apply the standard 1,900 / 3,900 Copilot allowances. Keep Sonnet 5.5 at $2/$10 for standard API tokens unless your contract says otherwise; model cache and task-level savings separately.
3. List every workload still on gpt-5.6-sol and schedule a GPT-6 Sol re-test for your next budget review; 21 November is a minimum supported-through date, not a confirmed price change.
4. Mark workloads that can move to cheaper IDs (GPT-6 Luna, Gemini 3.8 Flash, DeepSeek V4.1 Flash off-peak) without quality loss.
5. Put the Gemini 1 January rate step and Haiku 5.5 / generally available Gemini 4 availability in the finance calendar as review dates; no Haiku price or general Gemini 4 API price is confirmed.

Common mistakes

• Treating Anthropic's vendor-reported "up to 30% lower cost per task" for Sonnet 5.5 as a per-token rate reduction. Its listed input/output token rates remain $2/$10.
• Treating 21 November as a confirmed GPT-5.6 Sol price cliff, or assuming a post-promo rate before OpenAI announces one; compare GPT-6.1 Sol and GPT-6 Sol on your workload meanwhile.
• Using the flat list price as the forecast when your actual route is cached or batched. Caching and batch discounts change the effective rate, so model them before you commit a number.
• Assuming individual Copilot plans describe Business/Enterprise pools.
• Treating BYOK harness spend as "included in the IDE seat."

Bottom line

Review the GPT-5.6 Sol promo at 21 November, without assuming a price change, and diarise the confirmed Gemini 3.7/3.8 Flash rate step on 1 January. For Copilot, apply the current organization credit allowance shown in GitHub Docs and your admin view; this article does not assert a universal promotion-end date. Sonnet 5.5 is available at $2/$10 standard token rates; keep watching for Haiku 5.5 and general Gemini 4 availability without inventing prices. 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.

Sources (checked 2026-10-01 unless noted)

• OpenAI — Introducing GPT-6 Sol and Luna ($2/$10, 50% below GPT-5.6 Sol promo): https://openai.com/index/introducing-gpt-6-sol-and-luna/
• OpenAI — GPT-6.1 Sol announcement (2026-09-29): https://openai.com/index/introducing-gpt-6-1-sol/
• OpenAI — GPT-6.1 Sol model documentation (current base/cached prices and >272K price step): https://developers.openai.com/api/docs/models/gpt-6.1-sol — checked 2026-10-01
• OpenAI — GPT-5.6 Sol model documentation (current $4/$20 promotional price available at least through 2026-11-21; no post-promotion price stated): https://developers.openai.com/api/docs/models/gpt-5.6-sol — checked 2026-10-01
• OpenAI — GPT-5.6 launch and 2026-08-21 pricing update (historical context; does not state the current promotion end date): https://openai.com/index/gpt-5-6/
• Google — Gemini API pricing ($0.75/$3.75 through 2026-12-31, then $1.50/$7.50 for Gemini 3.7 and 3.8 Flash): https://ai.google.dev/gemini-api/docs/pricing — checked 2026-10-01
• Google — Gemini API deprecations (Gemini 3.7/3.8 shutdown status): https://ai.google.dev/gemini-api/docs/deprecations — checked 2026-10-01
• Anthropic — Claude API pricing (Sonnet 5.5 $2/$10; Opus 5.5 $4/$20): https://platform.claude.com/docs/en/about-claude/pricing
• Anthropic — Introducing Claude Sonnet 5.5 (released 28 September; same token rates as Sonnet 5): https://www.anthropic.com/claude-sonnet-5-5
• Anthropic — Introducing Claude Opus 5.5 (Haiku 5.5 announced for the coming weeks): https://www.anthropic.com/claude-opus-5-5
• Google — Gemini 4 Argon announcement (restricted Fairwind access): https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ — checked 2026-10-01
• DeepSeek — Models & Pricing (V4.1 Flash peak/off-peak rates and hours, including Chinese public holiday treatment): https://api-docs.deepseek.com/quick_start/pricing/ — checked 2026-10-01
• Meta — Muse Spark pricing and data-use terms by tier: https://dev.meta.ai/docs/pricing-rate-limits — checked 2026-10-01
• GitHub Docs — Plans for GitHub Copilot (Business $19 / 1,900 credits; Enterprise $39 / 3,900 credits): https://docs.github.com/en/copilot/get-started/plans
• GitHub Docs — Usage-based billing for organizations and enterprises (standard credits, pooled billing and budgets): https://docs.github.com/en/copilot/concepts/billing/organizations-and-enterprises/usage-based-billing
• GitHub Docs — Usage-based billing for individuals (credit allowances and legacy annual-plan exception): https://docs.github.com/en/copilot/concepts/billing/usage-based-billing-for-individuals
• GitHub Blog — Copilot transition to usage-based billing (announced 2026-04-27; transition date 2026-06-01): https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/