benchrHow this log works
Three rules, enforced since the first entry. Append-only: entries never get edited or removed, even when they age badly. That's what makes it a record. Same-day verification: a number enters the log only on a day benchr confirmed it on the provider's own pricing page or announcement, never from a search snippet or a third-party table. Source attached: every entry names where the number came from, so you can re-check it yourself.
That discipline has a cost: the log opens June 1, 2026, and refuses to backfill. A price archive you can trust beats a longer one you can't. Documented older moves appear below as context, with their own sources, clearly fenced off from the log.
The log so far
| Logged | Model | Event | Input | Output |
|---|---|---|---|---|
| 2026-06-10 | Claude Fable 5 | New model (Jun 9 release) | $10.00 | $50.00 |
| 2026-06-10 | GPT-5.4 | Baseline (Mar 5 release) | $2.50 | $15.00 |
| 2026-06-01 | Claude Opus 4.8 | Baseline (May 28 release) | $5.00 | $25.00 |
| 2026-06-01 | GPT-5.5 | Baseline (Apr 23 release) | $5.00 | $30.00 |
| 2026-06-01 | GPT-5 | Baseline (Aug 2025 release) | $1.25 | $10.00 |
| 2026-06-01 | GPT-5 Mini | Baseline | $0.25 | $2.00 |
| 2026-06-01 | Gemini 3.5 Flash | Baseline (May 19 GA) | $1.50 | $9.00 |
| 2026-06-01 | Grok 4.3 | Baseline (Apr release) | $1.25 | $2.50 |
| 2026-06-01 | DeepSeek V4-Pro | Baseline (Apr 24 release) | $0.435 | $0.87 |
| 2026-06-01 | DeepSeek V4-Flash | Baseline | $0.14 | $0.28 |
| 2026-06-01 | Kimi K2.6 | Baseline (Apr 20 release) | $0.95 | $4.00 |
| 2026-06-01 | Mistral Large 3 | Baseline (Dec 2 release) | $0.50 | $1.50 |
| 2026-06-01 | Llama 4 Maverick / Scout | Baseline (open weights) | $0 license | $0 license |
Most entries are baselines: the verified starting point each future change gets measured against. That's how an honest price database begins: you can't detect a change without a trustworthy "before."
Documented moves from before the log
These predate June 1, 2026, so they live outside the append-only dataset, but each is documented by the provider's own announcements and worth keeping in view:
The Opus lane fell two-thirds. Claude Opus 4 and 4.1 billed $15/$75 per million tokens through 2025. Since Opus 4.5 (November 2025), the lane bills $5/$25 — and both old models retire this summer, making the cut universal.
GPT-4o halved within months. The May 2024 launch snapshot billed $5/$15; later 2024 snapshots billed $2.50/$10. The original snapshot, still billing 2024 prices, shuts down October 23, 2026.
Reasoning got cheap. o1 launched December 2024 at $15/$60, and o1-pro hit $150/$600 in March 2025, OpenAI's priciest model ever. Today GPT-5.5 reasons better at $5/$30.
DeepSeek reset the floor twice. V4's April 2026 launch pricing undercut every closed model, and in June 2026 DeepSeek made its 75% V4-Pro reduction permanent at $0.435/$0.87, confirmed on api-docs.deepseek.com after third-party tables wrongly reported a rise.
And the increases
Headline rates rarely rise. Costs do — through routing and forced migration. Google's October 16 cutoff moves Gemini 2.5 Flash traffic onto a model billing 5× more per input token. xAI re-pointed legacy Grok API slugs to Grok 4.3 billing, repricing calls nobody touched. Gemini's over-200K-token surcharge doubled effective long-context rates for some workloads. A price database that only read rate cards would miss all three, which is why this log records sunsets and migrations alongside price changes.
What gets logged
Six event types, matching the dataset's schema: price_change, new_model, benchmark_update, deprecation_announced, sunset, and context_change. Pricing means published list rates for input, output, and (where offered) cached input; promotional and negotiated enterprise rates stay out. Free open-weight models are logged as $0-license entries so self-hosting comparisons stay honest about what "free" means. When a change lands, it also hits the site changelog and the affected pricing pages the same day.