Quesma released Quesma Shipper, an open-source collector that gathers AI agent session transcripts from tools like Claude Code, Codex, and Cursor into a user's own Amazon S3 bucket or similar storage.…
The Linux Foundation's Tokenomics Foundation, launched in August 2026, held its first event, Tokenomicon + FinOps X, in Amsterdam on September 22–23, gathering more than 500 FinOps and tokenomics prac…
Quesma hosted its first token economics hackathon, Session #5 of the Vibe Coding Summer Jam series, where five student teams analyzed 18 GB of coding-agent transcripts over one Friday evening to find …
GPT-6 Astra scored 63% on ARC-AGI-3 with the standard agent harness and 99% with a custom one, versus 8% for GPT-5.6 Sol and 30% for Claude Opus 5 on default harnesses, according to the ARC Prize blog…
JetBrains's SkillsBench benchmark found that RTK (Rust Token Killer), a terminal-output compression tool with over 79,000 GitHub stars, produced no consistent cost savings for AI coding agents, with D…
Gemini 3.8 Flash leads in mushroom recognition accuracy, but GPT-6 Astra demonstrates greater wisdom by asking for more photos and issuing disclaimers, according to a test by an unnamed researcher usi…
A Quesma test found that a single typical pull request consumed $270 of OpenAI Codex overflow usage within 24 hours on an OpenAI Business ChatGPT & Codex plan ($25/seat/month), exhausting the weekly l…
A developer's benchmark of Qwen3.8 27B quantizations found that the 17 GB Q4_K_M 4-bit version matches the full BF16 model on Terminal-Bench 2.1, while 1-bit versions collapse to near-random performan…
In mid-August, Gemini 3.7 Flash, Grok 4.6, GLM-5.3, and DeepSeek V4 Pro 0813 joined the frontier of LLM releases, and Quesma's Baba Is Bench rerun shows Gemini 3.7 Flash solving every trial at over 20…
Anthropic's Claude Code pricing in August 2026 shows the same tokens and model costing up to 40x more depending on whether users buy seats or pay per token, according to a report by Pragmatic Engineer…
Wiring four Claude models together in Claude Code to solve Terminal-Bench 2.1's 89 command-line tasks backfired in four ways, yielding a 78% solve rate and seventh place at a cost of $1,178, roughly t…
Wiring four Claude models together in Claude Code backfired in four ways on Terminal-Bench 2.1, costing $1,178 and solving 78% of 89 tasks for seventh place, roughly twice the cost of the top single-m…
An experiment by an unnamed researcher found that retrying failed requests in LLM benchmarks introduces selection bias, shortening average response lengths by 19.2% and reducing long-form outputs by u…
Claude Opus 5 achieved 100% pass@3 on the Baba Is You Intro stage, outperforming Kimi K3 (96%), Gemini 3.6 Flash (88%), and Grok 4.5 (75%) in a benchmark by Quesma, while also being more cost-effectiv…
A Quesma researcher burned through an entire Claude team x5 plan token limit in 30 minutes while researching AI agent tokenomics, prompting a redesign that uses multiple subscription services and chea…
Fable 5 and GPT-5.6 Sol can solve almost all levels in the first two stages of Baba Is You, but take significantly longer than humans, according to tests run by the developer of the open-source Baba I…
A new phenomenon called 'tokenflation' shows AI coding agents using excessive tool calls and reasoning for trivial inputs like 'Hi', costing developers time and money. Benchmarking 14 models revealed …
A developer benchmarked 14 AI agent models and found that the true cost of a simple greeting like "Hi" is not the token price (as low as $0.005) but the waiting time, which can cost up to $5.10 per gr…
Qwen 3.6 27B, a dense local language model from Alibaba's Qwen team, impresses developers with its general intelligence and practical coding abilities, running efficiently on consumer hardware via lla…