✅ Claude models are back to normal, and the earlier unavailability notice is withdrawn —— Anthropic’s status page
status.claude.com posted Resolved at 16:23 UTC on 3 September: the provider-side incident affecting claude-mythos-5-1, claude-fable-5-1, and claude-opus-5 ended at 00:16 on 4 September (UTC+8). The APIYI gateway was not changed and all claude-* models are serving normally again; Grok status is being tracked separately.📖 View details⚠️ Claude models are temporarily unavailable due to a provider-side incident, and Grok models are down at the same time —— Anthropic’s service is currently experiencing an incident, so requests to all
claude-* models fail or time out; follow the official status page status.claude.com for progress and the estimated recovery time. grok-* models are also interrupted and we are following up. Other routes: OpenAI models are running normally, with the OpenAI API official route largely normal and the Azure official route available as backup; Gemini models are unaffected. We will post here once service is restored.📖 View details💸
gpt-5.6-sol price cut synced: input $5 → $4, output $30 → $20, so the flagship tier is now cheaper than gpt-5.5 —— OpenAI lowered GPT-5.6 Sol pricing on 3 September (official $4 / $20, promotional rate guaranteed at least through 21 November 2026), and APIYI has synced: within 272K, $4 input / $20 output / $0.40 cached read, with cache writes at 1.25× the input rate ($5); above 272K the whole request is billed at 2× input and 1.5× output. Against the previous-generation gpt-5.5 ($5 / $30), Sol is 20% cheaper on input and a third cheaper on output; existing code migrates by changing only the model field.📖 View details📊 Google has published the official
gemini-3.8-flash spec sheet, confirming the 1M-token context —— The model page is up: a 1,048,576-token input limit, 65,536-token output limit, and text / image / video / audio / PDF input. Thinking offers only low / medium / high; minimal returns an error, matching our pre-launch tests. APIYI went live with the model on the evening of 2 September across both groups and both endpoints, so the earlier advice to wait for official specs before a full cutover no longer applies.📖 View details🚀
gemini-3.8-flash is live, and APIYI has it open for calls first —— Google’s newest Flash, out 2 September; its own model docs and launch blog have not listed this version yet. Pricing matches gemini-3.7-flash line for line ($0.75 in / $3.75 out per 1M tokens), so migrating means changing the model name and nothing else. 150 paired test cases across both protocols found capability parity with 3.7 and no regression unique to it.📖 View details🚀 iCover AI has been updated — SeeDance 2.5 is now selectable right in the model dropdown —— at
icover.ai/zh/seedance-official, SeeDance 2.5 (doubao-seedance-2-5-260628) now sits alongside 2.0, and all four task types (text-to-video, first frame, first/last frame, multimodal) plus aspect ratio, resolution and duration run with no code. For reference media, ingest into the asset library first and reference the asset:// ID: the request body shrinks to a few dozen bytes, create-task returns immediately, and asset IDs stay reusable across tasks for character consistency.📖 View details📖 On GPT-5.4+, tool calling with an explicit reasoning effort can be rejected on the chat endpoint — a request carrying
tools while explicitly sending a non-none reasoning_effort gets a 400: Function tools with reasoning_effort are not supported .... All four effort levels trigger it, omitting the parameter does not, and whether it fires depends on the upstream route — so “it worked last time” is not evidence you are safe. Move tool-carrying requests to /v1/responses, or set reasoning_effort="none". A new migration guide is up.📖 View details📊 When a Seedance call carries media, it is submission that is slow, not generation — your media travels upstream to APIYI, then on to Volcengine to be decoded and validated before the task ID comes back; inline Base64 stretches a one-second submission into tens of seconds, and one customer still got nothing at a 300-second read timeout. Ingest first and reference an
asset:// asset ID: the body shrinks to a few dozen bytes, the task ID returns immediately, and content checks move up to ingest time. A new how-to page is now live.📖 View details🚀
claude-fable-5-1 is live: cache reads fall from $1.00 to $0.25, and APIYI has matched the cut — Anthropic’s new Mythos-class flagship, shipped 1 September, alongside claude-fable-5-1-thinking. The cache read price is the headline change this generation; input $10 / output $50 per 1M tokens are unchanged, and our pricing matches the provider line for line. A 1M token context window, 128k max output, and availability across the default / svip / ClaudeCode groups on both endpoints. Check three breaking changes before migrating: forced tool use returns 400, thinking blocks are model-bound, and editing earlier turns invalidates them.📖 View details🗂️ Seedance 2.5 and the 2.0 family share the
SeeDance2 group (0.18x) — one token covers all four models — all four sit under one group for simpler management, with no separate token for 2.5 and no change to model name, endpoint or code. Measured rates come in two tiers: $12.60 per million tokens with no video in the input (480p / 720p / 1080p alike), and a lower $7.56 when the input contains video (multi-modal reference, video editing / extension).📖 View details📖 For earlier updates, visit the Live Updates Archive — browse history by month, category, or vendor.
Deep Dive
AI Radar
Telegram
Global users
Media Models
New additions