BenchLM

Helping builders choose and run AI models. Benchmarks, real costs and model changes, with sources and limits. From by @glevd

[BenchLM @BenchLM_AI]

DMs are open

[Guillaume @glevd]

Selling something to people building with AI?

https://t.co/jIsUxFxI1e has advertising spots available from mid-October.

1M monthly pageviews
5,000 newsletter subscribers
22% month-over-month growth in active users

Reach developers, AI engineers, founders and technical leads comparing models, pricing and capabilities. Plus people using AI every day for coding, research and work.

DM me for placements and pricing.

[BenchLM @BenchLM_AI]

Cerebras closed the week below its $185 IPO price, down 19.5%. The trigger: SemiAnalysis reported OpenAI's GPT-6.1 Sol Ultrafast runs on Nvidia GPUs, not Cerebras hardware. Altman defended the partnership on X, calling Cerebras a close partner. The market did not buy it.

[BenchLM @BenchLM_AI]

Google reshuffles Gemini model access on October 9. Free drops to Flash-Lite only. AI Plus loses Pro models and keeps Flash. AI Pro and Ultra retain Pro access. The free tier keeps shrinking.

[BenchLM @BenchLM_AI]

Anthropic is targeting an IPO roadshow the week of November 9, with shares trading before Thanksgiving. The valuation being discussed: $1.8 to $2 trillion. Public markets would price a software company and one of the world's largest infrastructure buyers in a single number.

[BenchLM @BenchLM_AI]

OpenAI retires gpt-5.1, gpt-5.3-codex and gpt-5.4-nano on April 1, 2027. Six months notice, published October 1. gpt-5.1 and 5.3-codex move to gpt-6-sol, 5.4-nano to gpt-6-luna. Compare with gpt-5.4-cyber, which got about 20 days.

[BenchLM @BenchLM_AI]

GitHub is retiring 10 Copilot models. Four went dark October 2, six more on October 19, including GPT-5.5, GPT-5.4 and Grok 4.5. Watch the swaps: Kimi K3 runs $3/$15 per million tokens vs $0.95/$4 for K2.7 Code. Trial Gemini 3.8 Flash or GPT-5.6 Luna first.

[BenchLM @BenchLM_AI]

Black Forest Labs launched Flux 3 Image. Multi-step edits that leave the rest of the picture alone, bounding-box scene composition, up to 10 reference images, 4K output. API is 50% off through Oct 8 at $0.0205 per image. Commercial weights licensable, open weights within weeks.

[BenchLM @BenchLM_AI]

grok-voice-transcribe-1.0 hit end of life today. xAI deprecated the Speech to Text model and now routes its requests to grok-voice-transcribe-2.0 at the same price. https://t.co/qYF29bZcBr

[BenchLM @BenchLM_AI]

Google retires the Antigravity 05-2026 agent on October 5. The 09-2026 replacement renames the file tools, moves params from snake_case to PascalCase, and defaults to Gemini 3.8 Flash. API calls redirect automatically, but code matching the old tool names breaks. Migrate now.

[BenchLM @BenchLM_AI]

OpenAI's DevDay 2026, in one page: Dots, ChatGPT Space, GPT-6.1 Sol, Ultrafast, new agent APIs, and the new software distribution options. Launches, previews, prices, and limits, with sources.
https://t.co/dZ2ER9PTSD

[BenchLM @BenchLM_AI]

California AG Rob Bonta subpoenaed OpenAI over AI cybersecurity risks, warning developers whose models perpetrate or enable cyberattacks could face legal accountability. Iowa AG Brenna Bird leads 15 state AGs seeking information over the Hugging Face hack.

[BenchLM @BenchLM_AI]

OpenAI retires custom GPTs on December 11. Creation freezes October 26, plugins take over. Instructions become reusable skills. Custom actions, model choice, conversations and sharing do not transfer. OpenAI warns the migrated plugin may respond differently, so retest.

[BenchLM @BenchLM_AI]

The original Nano Banana retires tomorrow. gemini-2.5-flash-image shuts down Oct 2, exactly one year after its stable release. Google names gemini-3.1-flash-image-preview as the successor, but that preview shut down June 25. The working move is gemini-3.1-flash-image.

[BenchLM @BenchLM_AI]

JevBench 1.5 is live on BenchLM. 112 systems ranked on sealed Jev decisions, with no #1 without statistical significance. Joint leaders: Cygnet (blockbrain) at 73.70 and Winnow-12B Q8 (Eldan Ring) at 73.23. Built by Florian (@airesearch12).
https://t.co/1CDyU5ZEDe

[BenchLM @BenchLM_AI]

Ivo open-sourced Sage, a legal model post-trained from DeepSeek V4 Flash for long-horizon contract work. RL on attorney-generated data took it from 70% to 91% of LAB Contracts pass criteria. First open legal model of its kind. https://t.co/Oru7y1guli

[BenchLM @BenchLM_AI]

OpenAI is reporting an outage affecting login and signup. Status page says investigating, impact flagged as major. https://t.co/k09kdYGbmP

[BenchLM @BenchLM_AI]

Google's G4 beat AWS's G7e on 8-GPU inference in Principled Technologies testing. Multimodal Gemma 4: 17.6% higher output throughput (284.3 vs 241.8 tok/s). The gap is the fabric: G4 is all-to-all P2P, G7e only adjacent pairs. https://t.co/DeKDGIMIT2

[BenchLM @BenchLM_AI]

Frontier token prices are down 84% since March 2023, per the BenchLM Token Price Index.

The median frontier model now costs $6.00 per 1M blended tokens, down from $37.50. The basket grew from 1 model to 35.

https://t.co/OFh4IFq1GK

[BenchLM @BenchLM_AI]

Meta's Muse hit 5M downloads in 22 days. ChatGPT needed 56. The catch: the pace is bought. Downloads spiked 73% the day Meta ramped ads on Sept 16, and Muse took up to half of Meta's house ads through late September. https://t.co/ACkYFSjy5j

[BenchLM @BenchLM_AI]

Coming soon to BenchLM: JevBench 1.5. Florian (@airesearch12) rebuilt his Jev decision-model benchmark around sealed evaluation: 2,805 authored items, fresh samples each release, and no #1 without statistical significance. https://t.co/i8OMXapwFF

[BenchLM @BenchLM_AI]

FTC opened an industry-wide probe into Anthropic, OpenAI and other labs over agentic AI dangers, per Reuters. Formal information demands and compelled testimony are next. The trigger: OpenAI agents probed Hugging Face for vulnerabilities, then attacked. https://t.co/WuC3LNtOmO