AI API Price Changes and Model Updates

OpenAI released GPT-6-Cyber on 2026-11-05 for authorized vulnerability investigation, reproduction and fix verification. Access requires Daybreak Red approval plus additional model approval, enabled per project. The pricing page and model page list only Standard short-context rates; Batch is not supported and no Flex or Fast rows exist. Not listed among the data-residency models, so no regional uplift is recorded. Model page: 1,050,000-token context window, 922,000 max input tokens, 128,000 max output tokens.

Bedrock model ID amazon.nova-2-lite-v1:0 (Global profile global.amazon.nova-2-lite-v1:0). Text, image and video input; text output; 1M-token context, up to 64K output tokens; optional reasoning. Launched December 2, 2025. Web grounding bills $0.03 per request. Available through Geo and Global cross-region profiles only (no single-Region endpoint in us-east-1).

Bedrock model ID amazon.nova-lite-v1:0. Text, image and video input; text output; 300K-token context, up to 10K output tokens. Launched December 5, 2024. No Global cross-region profile.

Bedrock model ID amazon.nova-micro-v1:0. Text-only model; 128K-token context, up to 10K output tokens. Launched December 5, 2024. No Global cross-region profile.

Bedrock model ID amazon.nova-pro-v1:0. Text, image and video input; text output; 300K-token context, up to 10K output tokens. Launched December 5, 2024. No Global cross-region profile.

Official Agent API model ID perplexity/sonar. Search-grounded answers with citations. Sonar Pro, Sonar Reasoning Pro and Sonar Deep Research are no longer listed after the September 27, 2026 Chat Completions retirement. Perplexity does not publish a context window or maximum output length for Sonar in the Agent API docs; unknown values are not represented as zero.

Official API model ID glm-4.7. Text model with a 200K-token context window and up to 128K output tokens. Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Official API model ID glm-5. Text model with a 200K-token context window and up to 128K output tokens. Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Official API model ID glm-5.2. Previous flagship, same base model as GLM-5.3 and the same price. 1M-token context, up to 128K output tokens. Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Official API model ID glm-5.3-flash. Low-cost multimodal model: text, image, video and file input; text output; 1M-token context, up to 128K output tokens. Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Official API model ID glm-5.3-flashx. Faster serving of GLM-5.3-Flash (about 200 tokens/s per Z.ai). Text, image, video and file input; 1M-token context, up to 128K output tokens. Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Official API model ID glm-5.3. Flagship model. Text input only, 1M-token context, up to 128K output tokens. Reasoning is always on (effort low, high or max). Z.ai does not publish a release date or Batch rate in its API pricing; unknown values are not represented as zero.

Effective date was not published by the provider in the source used for this event.

OpenAI released GPT-6 Rosalind on 2026-10-30 as a life sciences model. Access is limited to organizations approved for Rosalind Research through OpenAI's trusted-access program, for internal research by authorized users. The pricing page lists only Standard short- and long-context rates (no Batch, Flex, Fast or Priority rows); short context is 272K input tokens or fewer, per OpenAI's short/long context definition. Rosalind is not listed among the data-residency models, so no regional uplift is recorded. OpenAI publishes no model page or context window for it.

Official API model ID qwen-flash, currently equivalent to snapshot qwen-flash-2025-07-28. Text-only; 1,000,000-token context window, maximum input 997,952 tokens and maximum output 32,768 tokens. Pricing is tiered by input tokens per request: up to 256K, and 256K to 1M; the tier applies to every token in the request. The model page lists batch-file rates for Singapore but also marks batch inference as unsupported there, so no batch price is represented. Prices are USD list prices for the International (Singapore) scope; the official model page states that it shows original prices excluding limited-time promotions.

Official API model ID qwen3.7-plus, currently equivalent to snapshot qwen3.7-plus-2026-05-26. 1,000,000-token context window; maximum input 991,808 tokens (983,616 in thinking mode) and maximum output 131,072 tokens. Pricing is tiered by input tokens per request: up to 256K, and 256K to 1M; the tier applies to every token in the request. Rows store list prices. The official pricing page marks the qwen3.7-plus alias with a limited-time 20% discount and no published end date; the dated snapshot carries no discount. No International batch rate is published. Prices are USD list prices for the International (Singapore) scope; the official model page states that it shows original prices excluding limited-time promotions.

Official API model ID qwen3.8-flash. 1,000,000-token context window; maximum input 991,808 tokens (983,616 in thinking mode) and maximum output 131,072 tokens. One untiered rate covers thinking and non-thinking modes. Implicit cache hits and explicit cache reads cost $0.016 per 1M tokens; explicit cache creation costs $0.20 (5-minute TTL). Batch inference is not supported in the International scope. Prices are USD list prices for the International (Singapore) scope; the official model page states that it shows original prices excluding limited-time promotions.

Official API model ID qwen3.8-max (snapshot qwen3.8-max-0902, alias qwen3.8-max-2026-09-02). 1,000,000-token context window; maximum input 991,808 tokens (983,616 in thinking mode) and maximum output 131,072 tokens in both modes. One untiered rate covers thinking and non-thinking modes. Implicit cache hits cost $0.25 per 1M tokens; explicit cache creation costs $2.50 and explicit cache reads $0.17 (5-minute TTL). Batch inference is not supported in the International scope. Prices are USD list prices for the International (Singapore) scope; the official model page states that it shows original prices excluding limited-time promotions.

Official API model ID muse-spark-1.3, the latest Muse Spark version and Meta's recommendation for new work. Text, image, video, audio and PDF input; text output; 1,048,576-token context window. Muse Spark 1.2 and 1.1 remain available at the same Standard price and are not listed separately. Contributor-mode variants trade training use of your content for a lower price and are noted, not listed as separate rows. Availability varies by country under Meta's Geographic Use Policy (restricted territories excluded). Meta does not publish a maximum output length, Batch rate or release date in the API docs; unknown values are not represented as zero.

Official API model ID muse-wonder-1.0, in preview. Text, image, video, audio and PDF input; text output; 1,048,576-token context window; reasoning effort from minimal to max. Muse Wonder Full is available only in Muse Code subscriptions, not on Model API. Not currently available in South Korea. Availability varies by country under Meta's Geographic Use Policy (restricted territories excluded). Meta does not publish a maximum output length, Batch rate or release date in the API docs; unknown values are not represented as zero.

Released 2026-10-26 through the Responses and Chat Completions APIs. The official model contract has a 1,050,000-token context window and 128,000 max output tokens, with text and image input and text output. Official absolute input, cached-input, cache-write, and output rates are stored for Standard, Batch, Flex, and Fast processing across both sides of the strict greater-than 272,000 input-token boundary. Regional processing uses the official 10% uplift. Global processing plus US and EU data residency are supported, including Fast mode with EU data residency. Ultrafast is not offered for this model.

Official API model ID kimi-k3.1-pro-highspeed, the high-speed variant of Kimi K3.1 Pro released 2026-10-23, with the same capabilities at about 2-3x the output speed and a 1,048,576-token context window. Cache writes cost $4 per 1M tokens for the default 5-minute TTL and $8 for the 1-hour TTL; cache hits cost $0.20, the same as K3.1 Pro. The official docs do not publish a separate reasoning-token rate, a Batch rate, or a verified maximum completion limit; unknown values are not represented as zero.

Official API model ID kimi-k3.1-pro, the first K3.1-series model, released 2026-10-23. It has a 1,048,576-token context window, text/image/video input, tool calling, always-on reasoning, and low/high/max reasoning_effort with max as default. Cache writes cost $2 per 1M tokens for the default 5-minute TTL and $4 for the 1-hour TTL; cache hits cost $0.20. The official docs do not publish a separate reasoning-token rate, a K3.1 Pro Batch rate, or a verified maximum completion limit; unknown values are not represented as zero.

Effective date was not published by the provider in the source used for this event.

Generally available and stable since 2026-10-13. The 1,048,576-token unified context window supports a default synchronous maximum output of 65,536 tokens and Long Decode up to the full context limit. Only implicit caching is supported; explicit caching is not supported, and cache hits are priced 95% below input. Thinking tokens are billed as output. No Free Tier is available. Grounding is priced separately and is excluded from token prices. Introductory prices apply through 2026-11-30; higher prices take effect on 2026-12-01.

Active (latest), released 2026-10-07; max output 128000 tokens. Prompt <=100000 uses short prices; >100000 uses long prices for the whole request. All input tokens including cache reads and writes count toward the threshold. Batch input/output use published 50% discount; no Batch cache components asserted.

Stable, released 2026-10-06. Inputs: text, image, video, PDF; outputs: text and image. Max output 32768 tokens. Batch supported; caching, Flex and Priority unsupported. V1 output is text/thinking compatibility only; image output has a distinct structured price. Grounding fees are outside this token contract.

Public Preview. Launch pricing: 50% off for 2 weeks. Standard list prices retained separately. No official absolute promotion end date asserted; current launch quote is verified as of 2026-10-09T12:36:29Z. Open weights coming soon, not currently asserted as downloadable. No predecessor/successor relationship asserted.

Released 2026-09-29. The official model contract has a 1,050,000-token context window and 128,000 max output tokens. Official absolute input, cached-input, cache-write, and output rates are stored for Standard, Batch, Flex, and Fast processing across both sides of the strict greater-than 272,000 input-token boundary. Regional processing uses the official 10% uplift. US and EU data residency support Standard, Batch, and Flex; Fast is unavailable with EU data residency.

Released 2026-09-28 and listed as Active (latest), with retirement not sooner than 2027-09-28. Standard global Claude API pricing is $2 input, $0.20 cache read, $2.50 5-minute cache write, $4 1-hour cache write, and $10 output per 1M tokens. Batch input and output are 50% of Standard. The full 1M context window uses the same token rates. Regional and partner-cloud pricing differences are outside this global-base record.

Verified cross-model generation comparison; Fast is represented as introduced and never as an absent-to-zero price change.

High-throughput, low-latency, cost-efficient TTS model for conversational voice and production workloads. Stable Gemini API model released 2026-09-22 with no shutdown date announced. Official input and output token limits are 8,192 and 16,384. Supports Caching, Batch API, Flex inference, and Priority inference. Current prices are effective through 2026-12-31; separately stored future records take effect 2027-01-01. Audio output is 25 tokens per second. The legacy V1 output scalar remains null because the published output price is for audio tokens, not text tokens.

Flagship creative TTS model for studio-grade fidelity, expressive acting, accents, and stable long-form multi-turn speech. Stable Gemini API model released 2026-09-22 with no shutdown date announced. Official input and output token limits are 8,192 and 16,384. Supports Caching, Batch API, Flex inference, and Priority inference. Current prices are effective through 2026-12-31; separately stored future records take effect 2027-01-01. Audio output is 25 tokens per second. The legacy V1 output scalar remains null because the published output price is for audio tokens, not text tokens.

Verified cross-model generation comparison; the computed output reduction is derived independently from canonical prices.

Verified cross-model generation comparison; component deltas are derived from canonical prices and remain separate from the provider claim.

Released on 2026-09-21 and available on the public xAI API. Supports text and image input with text output; reasoning efforts are low, medium, high (default), and xhigh. Global Standard rates use the short tier through 200,000 total prompt tokens and the long tier above 200,000; cached prompt tokens count toward the threshold and the selected tier prices the whole request. Grok 4.7 Fast is twice the Standard token rates but is available only in Cursor and Grok Build, not on the public xAI API, so it is excluded from public pricing rows. Batch API is not supported, no separate cache-write price is published, and the US regional endpoint applies a 1.10x premium to input, cached input, and output tokens, including long-context rates.

DeepSeek V4.1 Flash was released on 2026-09-10. The official API model name is deepseek-flash and the pricing page identifies the served model version as DeepSeek-V4.1-Flash. V1 compatibility pricing uses the conservative Peak rate. Off-peak is 50% of Peak; Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday. The new rates took effect at 2026-09-10T04:00:00Z.

Effective date was not published by the provider in the source used for this event.

Effective date was not published by the provider in the source used for this event.

OpenAI released GPT-5.6 Cyber on 2026-08-07 as an alias for purpose-trained cybersecurity models, available to Daybreak Red customers once Daybreak Red is enabled for the requesting project. Responses API only; Batch is not supported and no Flex or Fast rows exist. Not listed among the data-residency models, so no regional uplift is recorded. Model page: 400,000-token context window, 272,000 max input tokens, 128,000 max output tokens.

Effective date was not published by the provider in the source used for this event.