Pricing
API Input List Price by Capability Tier and Vendor
Historical input-token list prices for separate OpenAI and Anthropic flagship and economy tiers. The chart compares vendor lines directly; it does not average the two vendors, because a median of two values is equivalent to their arithmetic mean.
API Input List Price by Capability Tier and Vendor
- OpenAI flagship
- Anthropic flagship
- OpenAI economy
- Anthropic economy
Anthropic flagship input pricing was $10 per million tokens in 2026-08-21.
Methodology
The chart uses one canonical measure: USD list price per one million uncached text input tokens for standard synchronous API processing and the vendor's base short-context band. Input and output tokens are not blended because their prices differ substantially and a blend would require a workload-specific input/output ratio. Cached input, cache writes, batch or flex discounts, priority or fast premiums, long-context uplifts, fine-tuning, regional processing and negotiated commitments are excluded. Historical per-1,000-token quotes are multiplied by 1,000; no currency conversion is needed because every included quote is in USD.
The vendor panel is fixed to OpenAI and Anthropic from 2023-07-11, the first date on which official material provides token-denominated input prices for both a high-end and a low-cost general-purpose model from both vendors. The panel is intentionally not expanded when later vendors publish prices: adding entrants only after they become cheap would create entry and survivorship bias. At each change point, each vendor observation is published as its own line. The fixed panel is transparent, not market-share-weighted, and does not represent an industry average.
Tier assignment is prospective and based on the vendor's contemporaneous product hierarchy, not on today's model names or retrospective benchmark scores. The flagship tier is the generally available standard model the vendor describes as its flagship, most capable standard model, or top broadly available family tier. Premium 'Pro' variants, restricted safety-lifted models, and models positioned for a specialized modality are excluded. The economy tier is the vendor's current-generation model explicitly described as its smallest, fastest, most affordable, nano, mini, Instant or Haiku tier. A cheaper superseded model does not remain in the economy series after the vendor launches a current-generation replacement, which avoids selecting legacy survivors solely because their sticker price is low.
Models enter on a dated official release or price announcement and leave when a replacement tier becomes available, access is officially suspended, or an official price change takes effect. This rule is visible in the brief June 2026 Fable 5 entry, suspension and redeployment rather than smoothing the outage away. When OpenAI's 2024-01-25 post announced a GPT-3.5 Turbo price for release 'next week' without an exact effective day, the publication timestamp is used and the uncertainty is stated in the point note. Temporary prices are retained as real list-price observations: GPT-5.6 Sol's 2026-08-21 promotional price is included and marked as guaranteed only through at least 2026-11-21.
Every change point links to the official page that caused the constituent change. The initial GPT-4 value is additionally backed by the Internet Archive's 2023-03-14 capture of OpenAI's launch page, and Anthropic's July 2023 values come from its dated official model-pricing PDF. Live vendor pages can be edited after publication, so dated launch posts and exact Wayback captures are preferred where available. Archive coverage is uneven, especially for interactive pricing tables; the series therefore omits unsupported historical prices rather than interpolating them, and starts in July 2023 rather than fabricating a January 2023 baseline.
The former two-vendor median was removed. Each parsed OpenAI and Anthropic input price is emitted as its own series; no vendor is averaged, imputed, or filled when its price is unavailable.
Frequently asked questions
Why does the chart use input price rather than a blended token price?
Input and output tokens are separately metered products with very different list prices. A blended figure would depend on an assumed workload ratio and cache-hit rate. Using uncached input price keeps the unit consistent through time and lets readers apply their own output usage separately.
What does capability tier mean here?
It is a position in each vendor's contemporaneous product family, not a fixed benchmark-score threshold. Flagship means the broadly available standard model the vendor identifies as its leading general-purpose tier; economy means the current-generation model it identifies as smallest, fastest or most affordable. The mapping is made at release time and is not rewritten using later labels.
Why are capability-tier API prices relevant to the AI industry?
Token list prices help set the direct inference cost faced by developers choosing hosted models. Tracking fixed vendor tiers shows how entry prices for leading and lower-cost offerings changed, but it does not measure total task cost, model efficiency, demand, or realized enterprise pricing.
Is this the average price of the whole model market?
No. It is a fixed panel of separate OpenAI and Anthropic lines. Fixing the panel makes the history comparable and avoids adding later entrants only after their low prices are known, but it does not represent vendor market share or every API provider.
Why can a tier price rise even when inference is generally getting cheaper?
The series follows current-generation tier replacements. A new economy or flagship model can cost more per token while delivering materially greater capability, so the line mixes within-model price cuts with product-generation upgrades. Point notes identify which mechanism caused each move.
Are free tiers, batch discounts and cached-token discounts included?
No. The benchmark uses paid, standard synchronous, uncached, short-context input list prices. Free quotas, batch or flex processing, cached input, long-context premiums, regional uplifts and contract discounts are excluded because they are not directly comparable across vendors and time.