Fleet Radar — 2026-09-07

_Synth: deepseek/deepseek-v4-pro · Sweep: perplexity/sonar-pro · 430 models live · +0 new / ~11 changed / -1 gone_

What changed today

OpenRouter catalog: 0 new models; 11 changed, all pricing or maxOut shifts.
Steepest repricings: qwen/qwen3-next-80b-a3b-thinking $0.15->$1.20/M, qwen/qwen3.5-397b-a17b $0.39->$2.34/M, deepseek/deepseek-v4-pro-0813 $1.049->$3.148/M, ~z-ai/glm-flash-latest $0.071->$0.238/M, ~z-ai/glm-latest $1.12->$3.52/M.
DeepSeek stable IDs doubled: deepseek/deepseek-v4-pro $0.955->$1.911/M; deepseek/deepseek-v4-flash $0.089->$0.177/M.
No catalog model gained new tool/structured-output/input-modality today; several only raised maxOut.

Google gemini-3.8-flash GA Sep 2: best reasoning/coding Flash at 3.7 Flash pricing ($0.75/$3.75 per M). Also gemini-3.8-flash-cyber launched, gated via Fairwind Program. Google blog Gemini API changelog
Gemini API added agentic video understanding Sep 1 for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Gemini API changelog
Anthropic released Claude Fable 5.1 GA and Mythos 5.1 restricted; 1M context, 128k max output, adaptive thinking, beta thinking/effort controls; prompt cache reads cut 75%; Sonnet 5 price hike cancelled, stays $2/$10. Anthropic release notes Axios
Claude Code weekly limits +25% permanently from Sep 14; current +50% temporary boost remains until then. ExplainX
OpenAI: no new model/pricing/API change surfaced in the last 48h. OpenAI release notes

Catalog: no features:+ additions; only maxOut increases on several GLM/DeepSeek/Qwen models.
Anthropic off-catalog: Fable 5.1 adds always-on adaptive thinking, beta thinking/effort, mid-conversation system messages, 1M context, 128k output; prompt cache reads -75%. Anthropic release notes
Google off-catalog: agentic video understanding in the Gemini API. Gemini API changelog

agentic-terminal-coding: codex (GPT-5.5 xhigh) remains terminal leader; claude remains strongest for hardest real-repo patch quality.
algorithmic-coding: ds (deepseek/deepseek-v4-pro) - fleet-annotated algorithmic/bulk reasoning; note 2x price rise.
hard-reasoning: agy/pro (Gemini Pro) - fleet-annotated hard reasoning + long context; claude is fallback for real-repo/orchestration depth.
long-context (>=512k): agy/pro (Gemini Pro) - fleet-annotated long-context tier; ds/dsf are 1M+ budget fallbacks after repricing.
vision / multimodal: agy/vision (Gemini 3.1 Pro) remains best-in-fleet image understanding.
image-generation: no fleet-native route surfaced today - gap.
embeddings: no fleet-native route surfaced today - gap.
cheap-bulk-generation: free* (NVIDIA Nemotron free tier) for zero-cost bulk; dsf as paid cheap fallback.
local / on-device: local/fm (Apple on-device) for instant/offline; local (Qwen3.5-9B MLX) for local codegen.
best-value (capability per dollar): free* for zero-cost throughput; dsf for paid value after DeepSeek price rise.

Pin ds to deepseek/deepseek-v4-pro and avoid deepseek/deepseek-v4-pro-0813 unless the 0813 vintage is required; 0813 is now $3.148/M vs $1.911/M.
Keep dsf on ~deepseek/deepseek-v4-flash-latest at $0.16/M; avoid deepseek/deepseek-v4-flash-0731 at $0.28/M unless the dated snapshot is required.
When available in agy, test agy/flash -> gemini-3.8-flash: Google claims better reasoning/coding at 3.7 Flash pricing.
If Claude Max exposes Fable 5.1, trial it on hardest claude orchestration jobs; 75% prompt-cache read cut and 1M context are material.

No benchmark data in catalog or sweep; rankings are inferred from fleet labels and vendor claims, not measured today.
Cannot verify Claude Fable/Mythos availability inside Claude Max, exact API aliases, or exact pricing beyond cited sources.
Cannot confirm gemini-3.8-flash or agentic video understanding is already live on Anoof's AI-Pro/agy route.
No image-generation or embedding model appears in fleet or today's changed catalog.
OpenAI changelog fetch only returned a publication timestamp, so a possible OpenAI update may be missed.

_No gaps: every source answered._

Fleet Radar · live data, never training · 2026-09-07