Loading benchmark data…
EyesInAI·Loading live benchmark data
We would like to count visits with Google Analytics, which sets a cookie. It is entirely optional — the site works exactly the same either way, and we never use it for advertising. What we set, and why.
Hosted models die on a schedule the vendor picks, announced by email, with a window usually measured in days — while the model id sits hard-coded in production. Nobody aggregates those emails publicly, so every team discovers each shutoff alone. This ledger publishes ours: every decommission notice our pipeline has received, the replacement the vendor pointed at, and — the number this page exists for — how much notice the email actually gave. The same ledger drives automation on our side: a deprecated id is never picked for a live routing path, and the bench stops testing a model once its date passes.
One notice’s recommended replacement was itself deprecated four weeks later; another “replacement” had ten days to live. Vendor advice optimizes for the vendor’s catalog, not your workload.
On our same-suite head-to-head, one dying model scored 94% and its recommended replacement 56%. The migration still had to happen — but blind trust in the replacement id would have shipped a regression. See the case study.
One end-of-life model collapsed to 11% on our suite in its final week while still serving. A deprecation date is not a guarantee of full quality until that date.
Announced Aug 15, 2026 · vendor’s replacement: deepseek-ai/DeepSeek-V4-Flash-0731
Two days of notice — the shortest window in this ledger. A production dependency on this id had one weekend to move.
DeepInfra deprecation email, 2026-08-15
Announced Aug 13, 2026 · vendor’s replacement: openai/gpt-oss-120b
The workhorse id of the free Groq tier. Groq gave a long runway from first notice (60 days) — but the final reminder landed 3 days out. Our own bench held the recommended replacement back from live routing after real-traffic scoring contradicted its synthetic pass — see the case study.
Groq decommission reminder, 2026-08-13 (first announced 2026-06-17)
Announced Aug 13, 2026 · vendor’s replacement: openai/gpt-oss-20b
Same story as its 70B sibling: long first notice, 3-day final reminder, and a replacement that passed the synthetic suite while failing real traffic.
Groq decommission reminder, 2026-08-13 (first announced 2026-06-17)
Announced Aug 6, 2026 · vendor’s replacement: Qwen/Qwen-3.5-9B
Together AI serverless deprecation notice, 2026-08-06
Announced Aug 6, 2026 · vendor’s replacement: moonshotai/Kimi-K3
Retired 13 days after Kimi K3 shipped — a same-family upgrade treadmill: the vendor points you at the successor, and the successor has a different cost and behavior profile.
Together AI serverless deprecation notice, 2026-08-06
Announced Aug 6, 2026 · vendor’s replacement: MiniMaxAI/MiniMax-M3
The recommended replacement is a DIFFERENT VENDOR’S model family — a reminder that a host’s "replacement" optimizes for the host’s catalog, not for your workload.
Together AI serverless deprecation notice, 2026-08-06
Announced Jul 16, 2026 · vendor’s replacement: Qwen/Qwen3.6-35B-A3B
Three days of notice.
DeepInfra deprecation email, 2026-07-16
Announced Jul 14, 2026 · vendor’s replacement: MiniMaxAI/MiniMax-M3
M2.7 was itself the recommended replacement for M2.5 ten days earlier on DeepInfra. Following vendor replacement advice put you on a model with 10 days to live.
Together AI deprecation notice + 1-week reminder, 2026-07-14 / 2026-07-21
Announced Jul 11, 2026 · vendor’s replacement: MiniMaxAI/MiniMax-M2.7
DeepInfra deprecation email, 2026-07-11
Announced Jul 11, 2026 · vendor’s replacement: google/gemma-4-31B-it
DeepInfra deprecation email, 2026-07-11
Announced Jul 11, 2026 · vendor’s replacement: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B
The recommended replacement here was ITSELF deprecated by Together AI four weeks later (see above). Replacement chains are not stable ground.
DeepInfra deprecation email, 2026-07-11
Announced Jul 9, 2026 · vendor’s replacement: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B
DeepInfra deprecation email, 2026-07-09
Announced Jul 9, 2026 · vendor’s replacement: meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo
Replaced by its own quantized variant at a different price point — the id changes, the invoice changes, the notice window was a week.
DeepInfra deprecation email, 2026-07-09
Announced Jul 3, 2026 · vendor’s replacement: Qwen/Qwen3.5-9B
The generous outlier: ~7 weeks of notice. Also cross-vendor replacement advice — a Google model retired in favor of a Qwen model.
Together AI serverless deprecation notices, 2026-07-03 and 2026-08-13
Announced Jul 3, 2026 · vendor’s replacement: MiniMaxAI/MiniMax-M3
Together AI serverless deprecation notice, 2026-07-03
Announced Jul 3, 2026 · vendor’s replacement: zai-org/GLM-5.2
Together AI serverless deprecation notice, 2026-07-03
Announced Jul 3, 2026 · vendor’s replacement: Qwen/Qwen3.5-9B
Together AI serverless deprecation notice, 2026-07-03
Announced Jun 17, 2026 · vendor’s replacement: openai/gpt-oss-120b
Sat in a live routing fast-path when the notice arrived. On our same-suite head-to-head the dying model scored 94% and its recommended replacement 56% — the replacement was worse than the deceased, and the migration still had to happen.
Groq deprecation email, 2026-06-17
Announced Jun 17, 2026 · vendor’s replacement: openai/gpt-oss-120b
By the final week it had collapsed to 11% on our expanded suite — end-of-life models can degrade in place before the shutoff date.
Groq deprecation email, 2026-06-17
The defense against week-long notice windows isn’t reading email faster — it’s never hard-coding a model id you can’t re-verify. Our routing picks models from a live benchmark, so a dying id drops out of the pool the day its notice lands, and its replacement has to earn the traffic.