TL;DR — This page tracks LLM API pricing the way it should be tracked: checked against official rate cards, with a running changelog of price changes — because in 2026 the story is not any single price, it is how fast they move. Last check: 2026-09-22. Cite freely; every number below carries an as-of date, and stale copies of this table elsewhere are exactly why we maintain it.
What was verified, and when
We would rather show our working than claim a blanket "all verified today," because a pricing table that overstates its own freshness is the problem it is meant to solve.
| Provider | Last verified | This round |
|---|---|---|
| Anthropic | 2026-09-13 | ✅ Re-verified — Sonnet 5 introductory rate confirmed permanent |
| OpenAI | 2026-09-13 | ✅ Re-verified — GPT-5.6 family added (we had missed it; cuts landed Jul 30) |
| 2026-09-13 | ✅ Re-verified — Gemini 3.8 Flash added with its Jan 1, 2027 step-up | |
| DeepSeek | 2026-09-14 | ✅ Re-verified against the official pricing docs — both rows confirmed unchanged |
| Zhipu | 2026-09-22 | ✅ Re-verified against docs.z.ai (USD list) and cross-checked on mistral.ai's marketplace table — GLM-5.2/5.1 corrected to the official $1.4/$4.4; GLM-5.3 family added |
| Moonshot | 2026-09-22 | ✅ Re-verified on platform.kimi.ai — Kimi K2.6 input rose $0.60 → $0.95; Kimi K3 and K2.7-code added |
| Mistral | 2026-09-22 | ✅ Re-verified on mistral.ai/pricing/api — flagship is now Medium 3.5 ($1.5/$7.5); Large 3 sits below it at $0.5/$1.5 |
The changelog below is the part we consider load-bearing: a price is only interesting next to what it used to be.
Current pricing (per million tokens, USD — see verification table above for per-provider dates)
| Provider | Model | Input | Output | Notes |
|---|---|---|---|---|
| Anthropic | Claude Fable 5 | $10.00 | $50.00 | Top tier |
| Anthropic | Claude Opus 5 | $5.00 | $25.00 | 3× cut vs Opus 4 |
| Anthropic | Claude Sonnet 5 | $2.00 | $10.00 | Introductory rate made permanent — the planned Sept 1 rise to $3/$15 was cancelled |
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 | Details |
| OpenAI | GPT-5.6 Sol | $4.00 | $20.00 | Promotional, at least through Nov 21, 2026 (was $5/$30) |
| OpenAI | GPT-5.6 Terra | $2.00 | $12.00 | −20% since Jul 30, 2026 (was $2.50/$15) |
| OpenAI | GPT-5.6 Luna | $0.20 | $1.20 | −80% since Jul 30, 2026 (was $1/$6) |
| OpenAI | GPT-5.5 | $5.00 | $30.00 | >272K-token prompts bill at 2× input / 1.5× output |
| OpenAI | GPT-5.4 | $2.50 | $15.00 | |
| OpenAI | GPT-4o | $2.50 | $10.00 | |
| OpenAI | GPT-4o-mini | $0.15 | $0.60 | Details |
| Gemini 3 Pro | $1.25 | $5.00 | ||
| Gemini 3 Flash | $0.075 | $0.30 | Cheapest input tracked | |
| Gemini 3.8 Flash | $0.75 | $3.75 | Introductory through Dec 31, 2026; rises to $1.50/$7.50 on Jan 1, 2027 | |
| DeepSeek | V4 Pro | $1.74 | $3.48 | Cache hits $0.014/1M; off-peak rates are half of peak |
| DeepSeek | V4 Flash | $0.14 | $0.28 | Off-peak rates are half of peak (peak = 01:00-04:00 & 06:00-10:00 UTC, Mon-Fri) |
| Zhipu (Z.ai) | GLM-5.3 | $1.40 | $4.40 | New; cached input $0.26 |
| Zhipu (Z.ai) | GLM-5.3-Flash | $0.15 | $0.50 | New; cached input $0.03 |
| Zhipu (Z.ai) | GLM-5.3-FlashX | $0.37 | $1.25 | New; cached input $0.075 |
| Zhipu (Z.ai) | GLM-5.2 | $1.40 | $4.40 | Official USD list (our earlier |
| Zhipu (Z.ai) | GLM-5.1 | $1.40 | $4.40 | Same as 5.2 on the USD list; cached input $0.26 |
| Zhipu (Z.ai) | GLM-4.7-Flash | Free | Free | 200K context; still free as of 2026-09-22 |
| Moonshot | Kimi K3 | $3.00 | $15.00 | New; 1M context; cache-hit input $0.30; cache write $3/$6 per 1M (5-min / 1-h TTL) |
| Moonshot | Kimi K2.7-code | $0.95 | $4.00 | New; cache-hit input $0.19; 262K context |
| Moonshot | Kimi K2.6 | $0.95 | $4.00 | Input +58% since our Aug 26 check (was $0.60); cache-hit input $0.16 |
| Mistral | Mistral Medium 3.5 | $1.50 | $7.50 | Current flagship ("long-horizon tasks, agentic coding") |
| Mistral | Mistral Large 3 | $0.50 | $1.50 | Priced below Medium 3.5 |
| Mistral | Mistral Small 4 | $0.15 | $0.60 | |
| Mistral | Codestral | $0.30 | $0.90 |
Cache discounts: Anthropic ~90% off cached input; Zhipu ~75–80% off; OpenAI cached-input discounts apply. Batch APIs: Anthropic −50% (stacks with cache); others vary. CNY→USD at ~7.15.
Price-change log — the part nobody else keeps
Dated, sourced changes only. This log is the citable asset: it turns "prices change fast" from a vibe into a record.
| Date | Change | Magnitude |
|---|---|---|
| 2026-09-22 | Kimi K2.6 input price $0.60 → $0.95 per 1M on the official rate card (output unchanged at $4.00); detected on our weekly re-check, exact effective date not published | +58% input |
| 2026-09-22 | GLM-5.2 / GLM-5.1 corrected to the official USD list, $1.4/$4.4 — our earlier |
Correction, +25% vs our prior figure |
| 2026-09 | New tiers: GLM-5.3 ($1.4/$4.4), GLM-5.3-Flash ($0.15/$0.50), GLM-5.3-FlashX ($0.37/$1.25); Kimi K3 ($3/$15, 1M context) and Kimi K2.7-code ($0.95/$4); Mistral's flagship is now Medium 3.5 ($1.5/$7.5) | New models |
| 2026-09-10 | DeepSeek released V4.1-Flash; we have not yet verified its rate card, so it is deliberately absent from the table above rather than estimated | New model |
| 2026-09 | Claude Sonnet 5 introductory pricing made permanent — the $2/$10 rate was due to rise to $3/$15 on Sept 1, 2026; Anthropic cancelled the increase | Increase cancelled |
| 2026-07-30 | GPT-5.6 family price cuts — Luna to $0.20/$1.20 (−80%), Terra to $2/$12 (−20%), Sol to $4/$20 promotional (−20% in, −33% out) | −20% to −80% |
| 2026 | Gemini 3.8 Flash introductory $0.75/$3.75 runs through Dec 31, 2026, then doubles to $1.50/$7.50 on Jan 1, 2027 | +100% scheduled |
| 2026-08 | Claude Opus 5 launches at $5/$25 — replacing Opus 4's $15/$75 as Anthropic's workhorse frontier tier | −67% frontier repricing |
| 2026-08 | Claude Sonnet 5 promotional pricing $2/$10 announced through Aug 31, 2026 | Promo |
| 2026-02 | GLM-5 family launches with 67–100% API price increases over prior generation (per launch coverage) | +67–100% |
| 2026 (ongoing) | Gemini 3 Flash holds the tracked floor for input pricing at $0.075/1M | Floor marker |
Have a dated change we missed? The log only accepts changes we can verify against official sources — that constraint is the point.
How to read this table like a router
Three structural facts matter more than any single price (full routing guide):
- The spread is the opportunity: ~200× between cheapest and most expensive input token. Routing exists because of this table.
- Repricing waves change routing math overnight — the Opus 5 cut moved the frontier threshold for every well-tuned router (analysis).
- The free tier is real: GLM-4.7-Flash at $0 with 200K context has no Western equivalent and quietly absorbs background workloads.
Methodology
- Verification cadence: every rate re-checked against the provider's official pricing page weekly (Fridays); the "verified" date above moves only when checks pass
- Sources: official rate cards only — no aggregator hearsay; CNY rates from open.bigmodel.cn with USD conversion noted
- Corrections: if you find a stale number, it is a bug — the whole point of this page is that stale pricing tables are everywhere and this one is not
- Per-model deep dives: Opus 5 · Haiku 4.5 · GPT-4o-mini · GLM family · market guide
- Head-to-head pages built on this table: the model comparison index — 24 pairs, each with monthly cost math and a routing verdict, updated whenever this tracker's numbers move
- Routing practice: offloading simple workloads to cheaper open-source models · routing + caching vs pricing and latency