What’s changed.

Provider billing condition · 2026-09-21

xAI changes X Search billing from calls to fetched results

Effective 21 September 2026 at 12:00 p.m. PT, xAI bills X Search at $5 per 1,000 posts fetched and $10 per 1,000 profiles fetched instead of $5 per 1,000 calls. The Grok task candidates and recommendations are unchanged.

Provider effective date: 2026-09-21. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: DAT-002 AGE-005

Evidence: docs.x.ai source

Provider watchlist · 2026-09-21

Gemini 3.8 Live enters the realtime watchlist

Google lists Gemini 3.8 Live as a stable Live API model with 131,072-token input context, separate text/audio pricing, and no caching or Batch support. It is watchlist evidence only; existing realtime task mappings are unchanged pending evaluation.

Provider effective date: 2026-09-15. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-007 REA-008

Evidence: ai.google.dev source · ai.google.dev source · ai.google.dev source

Provider watchlist · 2026-09-21

Google releases Antigravity Agent 09-2026 and schedules the prior preview shutdown

Google released the Antigravity Agent 09-2026 preview on 17 September and schedules the 05-2026 preview to shut down on 5 October. The new version is recorded for future coding, data and agentic evaluation; no task mapping or access policy changed.

Provider effective date: 2026-09-17. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: COD-003 DAT-003 AGE-001 AGE-004 AGE-005

Evidence: ai.google.dev source · ai.google.dev source · ai.google.dev source

Catalog correction · 2026-09-14

DeepSeek V4 Flash alias now serves V4.1 Flash at lower rates

DeepSeek now serves the legacy deepseek-v4-flash alias with DeepSeek-V4.1-Flash. Peak input, cached input and output rates are $0.30, $0.006 and $1.20 per 1M; off-peak rates are half. EDC retained the candidate label and task recommendation pending evaluation.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: DAT-005

Evidence: api-docs.deepseek.com source

Correction to the earlier item.

Provider availability · 2026-09-14

DeepSeek confirms V4 Pro API will continue

DeepSeek states that V4 Pro API service will continue after 14 September 2026 and that billing remains unchanged. This is an availability fact; the task recommendation remains pending evaluation.

Provider effective date: 2026-09-14. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-006

Evidence: api-docs.deepseek.com source

Provider retirement · 2026-09-14

Google schedules Gemini 3.1 Flash-Lite shutdown

Google schedules Gemini 3.1 Flash-Lite shutdown for 7 May 2027 and names Gemini 3.5 Flash-Lite as replacement. EDC did not change the content-moderation candidate without a task evaluation and working-fallback review.

Provider effective date: 2027-05-07. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CLA-002

Evidence: ai.google.dev source

Provider availability · 2026-09-14

Google names a replacement for Gemini 3 Flash Preview

Google names Gemini 3.6 Flash as the replacement for Gemini 3 Flash Preview but gives no shutdown date. EDC retained the helpdesk candidate mapping pending evaluation rather than treating the provider migration path as task-fit evidence.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: AGE-003

Evidence: ai.google.dev source

Source correction · 2026-09-14

DeepSeek retires the Vision Exp model behind its legacy alias

DeepSeek still accepts the deepseek-v4-flash-vision-exp alias, but the associated experimental model is retired; calls are now served by DeepSeek-V4.1-Flash and billed at Flash rates. This corrects the earlier watchlist item.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: api-docs.deepseek.com source

Correction to the earlier item.

Source correction · 2026-09-14

Gemini 3.8 watchlist item gains pricing evidence

Google lists Gemini 3.8 Flash at $0.75 input, $0.075 cached input and $3.75 output per 1M through 31 December 2026; Batch/Flex input and output are $0.375 and $1.875. It remains watchlist-only.

Provider effective date: 2026-09-02. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: ai.google.dev source · ai.google.dev source · ai.google.dev source

Correction to the earlier item.

Provider release · 2026-09-14

Step 3.7 Flash added to the provider watchlist

StepFun lists Step 3.7 Flash with 256K context, native image/video input, tool calling and direct API rates of $0.20 cache-miss input, $0.04 cache-hit input and $1.15 output per 1M. No task selection changed.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: platform.stepfun.ai source · platform.stepfun.ai source

Provider deprecation · 2026-09-07

GPT-Realtime Mini retirement is now scheduled

OpenAI schedules the gpt-realtime-mini alias to shut down on 20 January 2027 and names gpt-realtime-2.1-mini as its replacement. The active candidate remains unchanged pending task evaluation and fallback validation.

Provider effective date: 2027-01-20. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-008

Evidence: developers.openai.com source

Provider release · 2026-09-07

GPT-6 Astra added to the watchlist

OpenAI released GPT-6 Astra with a 1,050,000-token context window. Standard pricing below the long-context threshold is $10 input, $1 cached input, $12.50 cache write and $50 output per 1M tokens. It is watchlist-only; no candidate mapping or policy changed.

Provider effective date: 2026-09-03. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: developers.openai.com source · developers.openai.com source

Provider release · 2026-09-07

GPT-Realtime 2.1 family added to the watchlist

OpenAI released GPT-Realtime 2.1 and GPT-Realtime 2.1 Mini. The Mini pricing table lists text at $0.60/$0.06 cached/$2.40 and audio at $10/$0.30 cached/$20 per 1M tokens. The family is watchlist-only.

Provider effective date: 2026-07-06. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: developers.openai.com source · developers.openai.com source

Provider release · 2026-09-07

Claude Fable and Mythos 5.1 added to the watchlist

Anthropic released Claude Fable 5.1 and Mythos 5.1 with 1M context, $10 input, $0.25 cache read and $50 output per MTok, plus 50% Batch input/output rates. Fable is active; Mythos is invitation-only through Project Glasswing. No task recommendation changed.

Provider effective date: 2026-09-01. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: platform.claude.com source

Provider release · 2026-09-07

DeepSeek V4 Flash Vision Exp added to the watchlist

DeepSeek released DeepSeek-V4-Flash-Vision-Exp as an experimental multimodal API model and states that it matches V4 Flash on text capabilities while adding multimodal agent capabilities. It is watchlist-only.

Provider effective date: 2026-08-21. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: api-docs.deepseek.com source

Provider release · 2026-09-07

Gemini 3.8 Flash added to the watchlist

Google lists a 2 September 2026 release date for Gemini 3.8 Flash. It is recorded for a future task comparison; existing candidate mappings are unchanged.

Provider effective date: 2026-09-02. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: Watchlist; no candidate mapping changed.

Evidence: ai.google.dev source

Billing clarification · 2026-09-07

Voice tasks now state the billing unit

The two voice-task rows retain their text-token prices, explicitly labeled. Realtime 2 audio input/output are $32/$64 per 1M audio tokens. Mini audio pricing requires a further source check for the exact model. These rows are not full voice-session estimates.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-007 REA-008

Evidence: developers.openai.com source · developers.openai.com source

Catalog correction · 2026-09-07

GPT-5.6 Sol — catalog values refreshed

Direct OpenAI API Standard, short-context tier. Long-context input / output / cached input: $8 / 30 / 0.8 per 1M. Cache writes, regional processing and tools can cost extra. Batch/Flex rates differ. Sol promotional pricing is available at least through 21 November 2026.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: COD-003 REA-002

Evidence: developers.openai.com source

Catalog correction · 2026-09-07

GPT-5.4 mini — catalog values refreshed

Direct API text-token pricing. Cache value is cached input; previous catalog rounded it to $0.07.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: COD-004

Evidence: developers.openai.com source

Catalog correction · 2026-09-07

Gemini 3.1 Flash-Lite — catalog values refreshed

Gemini Developer API paid Standard. Text/image/video input rates; audio input may differ. Output includes thinking tokens. Cache read price excludes $1 per 1M tokens per hour storage and grounding/tool charges.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CLA-002

Evidence: ai.google.dev source

Catalog correction · 2026-09-07

Gemini 3 Flash Preview — catalog values refreshed

Gemini Developer API paid Standard. Text/image/video input rates; audio input may differ. Output includes thinking tokens. Cache read price excludes $1 per 1M tokens per hour storage and grounding/tool charges. Serving model is gemini-3-flash-preview; no shutdown date is announced on the provider deprecation page.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: AGE-003

Evidence: ai.google.dev source · ai.google.dev source

Catalog correction · 2026-09-07

Qwen3.7 Flash — catalog values refreshed

Alibaba Model Studio International, Singapore endpoint, input ≤32K. Cache price is implicit caching; explicit cache reads $0.003 and writes $0.038 per 1M. Longer prompts and other regions use different tiers. Batch is unavailable in this serving region.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CON-006

Evidence: www.alibabacloud.com source

Catalog correction · 2026-09-07

Kimi K3 — catalog values refreshed

Direct Kimi API rates, excluding taxes. Cache price is automatic cache-hit input; output includes reasoning. Context 1,048,576 tokens.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: COD-005

Evidence: www.kimi.ai source

Catalog correction · 2026-09-07

DeepSeek V4 Flash — catalog values refreshed

Direct DeepSeek API PEAK rates, Monday–Friday 01:00–04:00 and 06:00–10:00 UTC. Off-peak input / output / cache hit: $0.22 / 0.66 / 0.007 per 1M. This replaces the earlier single-price assumption.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: DAT-005

Evidence: api-docs.deepseek.com source

Catalog correction · 2026-09-07

DeepSeek V4 Pro — catalog values refreshed

Direct DeepSeek API PEAK rates, Monday–Friday 01:00–04:00 and 06:00–10:00 UTC. Off-peak input / output / cache hit: $0.66 / 1.98 / 0.022 per 1M. This replaces the earlier single-price assumption.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-006

Evidence: api-docs.deepseek.com source

Catalog correction · 2026-09-07

Holo3.1-35B-A3B — catalog values refreshed

H Company direct API, holo3-1-35b-a3b. Cache pricing is not stated in this source. A missing cache price is not zero.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: VIS-004

Evidence: hcompany.ai source

Catalog correction · 2026-09-07

Mistral Small 4 — catalog values refreshed

Mistral direct inference API token rates. Cache price is cached input; verify deployment and service options separately.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CLA-003

Evidence: docs.mistral.ai source

Catalog correction · 2026-09-07

Ministral 3 3B — catalog values refreshed

Mistral direct inference API token rates. Cache price is cached input; self-hosting has separate infrastructure costs.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CLA-004

Evidence: docs.mistral.ai source

Catalog correction · 2026-09-07

Step 3.5 Flash — catalog values refreshed

StepFun direct API rates for step-3.5-flash. Cache price is cache-hit input; step-3.5-flash-2603 currently has the same listed rates.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: CLA-005

Evidence: platform.stepfun.ai source

Catalog correction · 2026-09-07

GPT Realtime 2 — catalog values refreshed

TEXT tokens only: $4 input / $24 output / $0.40 cached input. Audio tokens instead cost $32 / $64 / $0.40 per 1M. This row is not a complete voice-session cost estimate. Image tokens have separate rates.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-007

Evidence: developers.openai.com source

Catalog correction · 2026-09-07

GPT Realtime mini — catalog values refreshed

TEXT tokens only. Audio pricing was not confirmed for this exact model during this check; do not use the text price as a voice-session estimate. The page flags the dated 2025-12-15 snapshot as deprecated; verify the deployment before adopting.

Provider effective date: Not established by this check. Detected dates describe our check, not necessarily a new provider announcement.

Affected tasks: REA-008

Evidence: developers.openai.com source