Why one window, three providers, and a stale comparison
A cost estimate for calling GPT, Claude, or Gemini is really two numbers multiplied together: how many tokens your prompt turns into, and what each token costs. Most comparisons only track the second number, because it is the one that shows up on a pricing page. Between April and July 2026, OpenAI, Anthropic, and Google each shipped an update that moved one or both numbers, independently of each other and without any shared timing.
That is what makes this a reset rather than three unrelated announcements. A cost comparison written in early 2026, before any of these three updates shipped, is now wrong on two separate axes at once for every provider it covers. This hub exists to lay out exactly what moved, provider by provider, sourced and dated, so the comparison does not have to be taken on faith.
Structural Comparison Matrix
| Provider | Token-count change | Price change | Shipped | How confirmed |
|---|---|---|---|---|
| OpenAI GPT-5.6 | Near-1M context window (up to 922K input, 128K output) on the o200k_base encoding | Prompt caching moved from free and implicit to explicit breakpoints, with a 1.25x premium on cache writes | GA 2026-07-09 | Paraphrased, aggregated from model-tracking sources, not OpenAI’s own announcement post directly |
| Anthropic Claude Opus 4.8 | Tokenizer inherited from Opus 4.7 counts up to ~35% more tokens for the same text than pre-4.7 models | Base price unchanged ($5/$25 per MTok); Fast Mode cut from $30/$150 to $10/$50 per MTok | 2026-05-28 | Price paraphrased (third-party pricing coverage); tokenizer claim paraphrased, not in Anthropic’s own official GA changelog |
| Google Gemini 3.6 Flash | Uses 17% fewer output tokens than Gemini 3.5 Flash for comparable tasks | $0.75 / $3.75 per million input/output tokens now (through 2026-12-31), rising to $1.50 / $7.50 in 2027, 1M-token context window | 2026-07-21 | Paraphrased (token reduction) / confirmed on Google’s own pricing page (both price points) |
One number in this table is confirmed against a primary source rather than paraphrased from secondary coverage: OpenAI’s own deprecations page states a minimum six-month retirement notice for generally available models, and separately lists gpt-4, o1, and o4-mini shutting down October 23, 2026, with gpt-5.6-sol and gpt-5.6-terra as the named replacements. Everything else above is labeled honestly as paraphrased, because it’s aggregated from third-party tracking and reporting, not each provider’s own primary documentation.
What changed, provider by provider
OpenAI widened GPT-5.6’s context window to roughly 1.05M tokens, a figure that matches OpenAI’s own model documentation, and switched to the o200k_base encoding, the same one GPT-4o introduced, not the older cl100k_base encoding GPT-3.5 and GPT-4 used. It also ended free, automatic prompt caching in favor of explicit cache breakpoints, a 1.25x premium on cache writes, and a 30-minute minimum cache lifetime, while keeping the roughly 90% discount on cache reads. Underneath that, the older gpt-4, o1, and o4-mini model IDs stop responding entirely on October 23, 2026, a hard cutover, not a soft warning.
Anthropic kept Claude Opus 4.8’s base price identical to Opus 4.7 ($5 per million input tokens, $25 per million output tokens, both confirmed on Anthropic’s own pricing page) while cutting Fast Mode from $30/$150 to $10/$50 per million tokens. The harder number to pin down is the tokenizer: third-party technical coverage reports that the tokenizer Opus 4.7 introduced, and that 4.8 inherited unchanged, counts up to roughly 35% more tokens for the same input text than the tokenizer pre-4.7 models used. Anthropic’s own official GA changelog for Opus 4.7 does not mention a tokenizer change at all, so treat that figure as reported, not confirmed.
Google launched Gemini 3.6 Flash at $0.75 per million input tokens and $3.75 per million output tokens, confirmed directly on Google’s own Gemini API pricing page, with a 1M-token context window, and reports it uses 17% fewer output tokens than Gemini 3.5 Flash for comparable tasks. That same page states the rate rises to $1.50 input / $7.50 output on 2027-01-01 — a scheduled increase, not the current price, worth flagging separately since an earlier version of the linked post below stated the future rate as if it were already in effect. Two cost levers are moving at once here regardless of which date applies: a lower per-token rate, and fewer tokens spent per task, which a sticker-price-only comparison misses entirely.
The 10 pieces of this reset
This hub is the starting point for a 10-part series. All ten pieces are published below.
- GPT-5.6’s Context Window: 922K In, 128K Out Tokens: the new near-1M window and the
o200k_baseencoding switch. - GPT-5.6 Ends Free Prompt-Cache Writes (1.25x Premium): explicit cache breakpoints replace free, automatic caching.
- OpenAI Extends Prompt Cache Retention to 24 Hours: a much longer window for repeated-call workflows to hit the cache.
- gpt-4, o1, and o4-mini Shut Down October 23, 2026: the one primary-sourced date in this series, straight from OpenAI’s own deprecations page.
- Claude Opus 4.8: Same Price, Cheaper Fast Mode: base pricing unchanged, Fast Mode cut roughly in half.
- Claude’s New Tokenizer Counts Up to 35% More: a reported change Anthropic’s own GA changelog doesn’t confirm.
- Claude Opus 4.7 Breaks temperature and top_p Params: a reported breaking change for a common sampling-parameter combination.
- Gemini 3.6 Flash Cuts Output Tokens 17%, Price Too: two cost levers moving independently, not one.
- OpenAI vs Claude: Prompt Caching Cost Math in 2026: why a single discount percentage no longer describes either provider’s caching economics.
- Same Prompt, Different Bill: GPT-5.6 vs Claude vs Gemini: the closing synthesis, tying every provider’s change into one cost picture.
Check your own numbers before you budget
None of the figures above should be the last word on your own costs. Token counts depend on your actual prompts, and every provider’s own API response carries the real, billed count in its usage (or, for Gemini, usageMetadata) field, more reliable than any estimate in this post. If you want a fast check before committing to a call, this site’s AI Token Counter gives an exact GPT count next to clearly labeled Claude and Gemini estimates, side by side, entirely in your browser, with no API key required — and the LLM Pricing Calculator turns those counts into an actual side-by-side dollar comparison across all three providers, cache math included.
Treat this hub, and every post in this series, as a snapshot with a real date attached, not a permanent reference. Re-check pricing and token counts against each provider’s current documentation before a budget review, not just once when a model first ships.
Browse more coverage like this in the AI Productivity archive.
