33 events in a year. Then 155 in the next.
Every model-market event Vitna has recorded since Mar 2023, newest first. Each entry links to a primary source on a domain the provider itself publishes on, and carries the date that source was last fetched; the newest of those checks is 2026-11-19. The log only ever gains entries.
Every event at its true date, newest left.
Each move as a share of the price before it.
83 detected moves carry both of their periods, so each bar is the move as a share of the price before it: cuts run left, rises run right, and the longest bar is the largest move. Each model leads to its own page and each meter to the move's own page. Curated entries cite an announcement rather than two readings and list below without a bar.
Every entry, with the document behind it.
10 of 216 events. policy.
- 2026Anthropicpolicysource ↗
Claude Opus 4.1 retired
Corrected. The row that supersedes this one stands in the record.
Deprecated June 5 with 60 days' notice; requests to the model now fail. The recommended replacement is Opus 4.8.
- Anthropicpolicysource ↗
US export-control directive suspends Fable 5 and Mythos 5
All customer access disabled three days after launch; restored at the end of June.
- Googlepolicysource ↗
Flex and Priority inference tiers open on the Gemini API
Two service levels at two prices for the same models, the latency-for-money trade the batch discount makes in reverse.
- Anthropicpolicysource ↗
The 1M token context window leaves beta at standard pricing
Requests over 200k tokens work with no beta header on Opus 4.6 and Sonnet 4.6, and are billed at the same per-token rate as a short one.
- Anthropicpolicysource ↗
US-only inference is priced at a premium
Data residency controls arrive with the inference_geo parameter; pinning inference to the US costs 1.1x standard rates on models released after 1 February 2026.
- 2025Anthropicpolicysource ↗
Global endpoint pricing introduced on Bedrock and Vertex AI
Regional routing starts carrying a premium over global on the partner clouds. First-party Claude API pricing is unaffected.
- 2024Anthropicpolicysource ↗
The Message Batches API opens at half the standard rate
Large asynchronous batches are processed for 50 percent less, matching the discount OpenAI set in April.
- OpenAIpolicysource ↗
Prompt caching, distillation and the Realtime API arrive together
Caching lands on the same day as the Realtime API, model distillation, image fine-tuning and Evals.
- Anthropicpolicysource ↗
Prompt caching enters beta on the Claude API
A cached prompt prefix is re-read at a fraction of the input rate, which changes the unit price of a repeated system prompt rather than the model.
- OpenAIpolicysource ↗
The Batch API arrives and halves the rate for patient work
Asynchronous request processing at a discount to the synchronous rate, the first time latency is priced separately from capability.
An event appears here only with a primary-source URL verified on the date the entry carries. Secondary coverage can corroborate a date but never stands as the source. A detected price move is published only after a second read at least an hour later agrees, and cites the archived document either side of it.
A week with 3 events or more is a page of its own, with its counts, its sized moves and every entry in it. The newest is the week of 16 Nov 2026, with 16 price moves, and all 16 are listed by week.
This log is published as a JSON Feed at /api/v1/feed.json, public and with no key, carrying every entry here and the decisions the engine has proposed against them, and as RSS at /log/rss.xml for a feed reader, with the same items under the same ids. Every headline above is also a page of its own, at an address that keeps pointing at it. Nothing in either feed is a recommendation to act.
