Last substantively updated: 30 October 2026. Quarterly review cycle: next full review due 30 January 2027. Known price changes and model retirements are checked sooner.
AI model costs have changed materially since this tracker’s July baseline. The useful comparison is now between specific models, service tiers and workloads, rather than a single US-versus-China discount. This update replaces the old API comparison with a dated sample of first-party rate cards and separates hosted API pricing from the economics of running open-weight models.
What do AI model APIs cost in October 2026?
The table is a selected comparison, not an exhaustive model directory or a capability ranking. Prices are in US dollars per million tokens. Input means uncached input unless stated otherwise. The comparison excludes batch and flex discounts, cache writes and reads, taxes, paid tools and regional-processing premiums. Provider conditions differ: context length, deployment location and promotional periods are shown where relevant.
| Model | Provider | Input / 1M tokens | Output / 1M tokens | Conditions |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | $10.00 | $50.00 | Standard, short-context rate |
| GPT-6.1 Sol | OpenAI | $2.00 | $10.00 | Standard, short-context rate |
| GPT-6.1 Terra | OpenAI | $1.00 | $5.00 | Standard, short-context rate |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | Standard, short-context rate |
| Claude Fable 5.1 | Anthropic | $10.00 | $50.00 | Base token rates |
| Claude Opus 5.5 | Anthropic | $4.00 | $20.00 | Base token rates |
| Claude Sonnet 5.5 | Anthropic | $2.00 | $10.00 | Base token rates |
| Claude Haiku 5.5 | Anthropic | $0.10 | $0.50 | Prompts up to 100,000 tokens |
| Gemini 4 Argon | $2.00 | $10.00 | Promotional rates through 30 November 2026 | |
| DeepSeek V4.1 Flash | DeepSeek | $0.15 / $0.30 | $0.60 / $1.20 | Off-peak / peak; uncached input |
| DeepSeek V4 Pro | DeepSeek | $0.66 / $1.32 | $1.98 / $3.96 | Off-peak / peak; uncached input |
| Qwen3.8 Max | Alibaba Cloud | $2.00 | $6.00 | Singapore, International deployment, up to 1M input |
| GLM-5.3 | Z.ai | $1.40 | $4.40 | Standard API rates |
| GLM-5.3 Flash | Z.ai | $0.15 | $0.50 | Standard API rates |
Sources checked 30 October 2026: OpenAI API pricing, Anthropic pricing, Google Gemini pricing, DeepSeek pricing, Alibaba Cloud Model Studio pricing and Z.ai pricing.
The current entry tiers overlap across countries of origin. GPT-6 Luna and short-prompt Claude Haiku 5.5 both list $0.10 input and $0.50 output per million tokens. DeepSeek Flash’s off-peak rates are $0.15/$0.60, rising to $0.30/$1.20 at peak. These prices alone do not establish equivalent quality, latency or reliability. A blanket claim that Chinese models are always five to ten times cheaper is no longer a useful buying rule.
Price changes and conditions to watch
Google lists Gemini 4 Argon at $2 input and $10 output per million tokens through 30 November 2026, increasing to $4/$20 from 1 December. That scheduled change should be reflected in this tracker before the next quarterly review. Its output charge includes thinking tokens.
DeepSeek’s peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. All other times are off-peak. The deepseek-flash API name currently serves DeepSeek-V4.1-Flash-1017; older V4 Flash names route to that model. Search and Fetch are separate charges, with per-request billing scheduled to begin on 1 November 2026.
Alibaba’s row uses its Singapore International deployment. The same model’s China and Hong Kong rates differ, so a lower regional quote should not be substituted without checking deployment and data-handling requirements. Claude Haiku 5.5 also has a different rate above 100,000 prompt tokens. OpenAI’s table distinguishes processing modes and context tiers; the rows above use Standard short-context rates.
Kimi is not included in the numeric table because this review could not reliably extract its complete current rate table. Its official pricing documentation does flag a lifecycle change: K2 models become unavailable to newly registered users from 31 October 2026, with full retirement planned by the end of December. Check the current K3 model and cache charges directly before budgeting.
Compare cost per successful task, not just token prices
For a simple uncached request, token cost equals input tokens multiplied by the input rate, plus output tokens multiplied by the output rate. Divide each token count by one million when using the table. Then add tool calls, storage and any other separately billed services.
For example, 1 million input tokens plus 200,000 output tokens costs $4 at a $2/$10 rate. The same token quantities cost $3.20 at Qwen3.8 Max’s Singapore International $2/$6 rate. Keep the workload fixed when comparing providers.
What changes the real bill?
Reasoning tokens, retries, conversation history, agent loops and failed outputs can materially change usage. Cache discounts matter for repeated prompts, while cache-write fees and retention periods differ. Measure accepted outputs on the same test set and count human review and rework. A lower price per token can still produce a higher cost per completed task. No universal token-verbosity multiplier is assumed here.
Are open-weight models free?
Downloadable weights and a hosted API are different products. A model’s licence governs whether and how its weights may be used; an API rate card prices a provider’s service. Self-hosting adds accelerator rental or depreciation, memory, utilisation, engineering and operational costs. Verify the licence for the exact release rather than assuming that every Chinese model is open source or commercially unrestricted. See our guide for builders using Chinese and open-source models.
API pricing, subscriptions and training costs are separate
Consumer chat subscriptions do not represent an API token allowance unless the provider explicitly says so. Local prices, promotions, taxes and product limits can change independently of API rates. For availability by country, use our AI model access guide, checking its own review date.
Training-run estimates and company-wide infrastructure spending also answer different questions from inference pricing. This refresh removes the old point-in-time spending and benchmark comparisons from the current buying guidance rather than treating July’s estimates as October facts. For model capability comparisons, consult the separately maintained capability tracker.
Update log
30 October 2026: Replaced the July API table with first-party rates for selected current models; corrected the blanket US–China discount claim; added deployment, peak-hour and promotional conditions; removed stale subscription, benchmark and capex claims from current guidance. Next full review: 30 January 2027. Earlier checks: DeepSeek tool billing on 1 November, Gemini promotional expiry on 1 December and Kimi K2 retirement by year-end.
July 2026: Original tracker baseline. Earlier rates describe that historical period and should not be used as current quotes.
