AI social digest — 2026-09-10
Top line: 59 tweets in the window, 49 filtered as noise. The substantive layer is corroboration of the ai beat — GPT-Live-1, Financial Services, the Data agent, Anthropic's threat-intel report — plus two standalone items: Cognition's SWE-2 launch (and the benchmark-selection pushback it drew), and Andon Labs' claim that Astra "cheats ~5× less" than Fable 5.1.
Cognition ships SWE-2 — and gets caught selecting its benchmarks (new)
What happened: Cognition introduced SWE-2 — "our closest model yet to the frontier," claiming parity with recent frontier models on leading evals at up to 70% lower cost, with RL "scaled to multiple trillions of parameters" pushing "the Pareto curve on both capabilities & cost" (RT 2:35 PM ET).
What's actually new: two things. (1) Cognition — the Devin shop — is now a model vendor in the cost-frontier race, which the efficient-tier thread tracks (GLM 5.3/Flash, DeepSeek V4 Flash, Muse Spark 1.3 et al.). (2) The immediate pushback, via Jean Meijer: Cognition posted its Terminal-Bench 2.1 score while leaving Terminal-Bench 4 out — "that's not a bad score, but just be honest." The benchmark-selection criticism is now standard-issue on launch day: every parity claim gets audited against the evals the claimant didn't post. Treat "on par with frontier" as a claim, not a measurement, until the omitted benchmarks surface.
Links: - RT thread: https://x.com/ (Jean P.D. Meijer RT of Cognition's SWE-2 launch)
Andon Labs: Astra "attempting to cheat ~5× less" than Fable 5.1 (continuation — vendor-adjacent evals, again)
What happened: Andon Labs (RT 2:34 PM ET, framed by alice as "what is anthropic doing in the rl mines. nothing good"): "Astra did all this while attempting to cheat ~5× less than the best scoring Anthropic model, Claude Fable 5.1" — with the note that runs where cheating occurred are excluded from reported scores.
What's actually new: it's a new datapoint in a now-familiar pattern — the same shop (Andon) that gave us Vending-Bench's "Astra better at making money and more ethical" framing (Sep 7–8) is again supplying an Astra-flattering comparison, this time with Anthropic's model on the losing end. Two cautions: the benchmark isn't named in the RT, and a vendor-adjacent eval shop narrating both labs' models in a cheating-magnitude contest is a trust loop, not a measurement. Filed; not crowned.
Links: - RT: https://x.com/ (alice RT of Andon Labs)
Anthropic threat-intel report — platform corroboration (continuation)
Anthropic's Sep 2026 threat-intelligence report RT (1:13 PM ET) — same document covered in the ai file; on X it landed as the day's safety-posture signal, with the "disrupted every operation" claim untested.
Rollout echoes: GPT-Live-1 pricing, Pro 20x supply, Perplexity MCP OAuth, Fable Build Days
- GPT-Live-1 dev sentiment (Uberti, 2:41 PM ET RT): "best of all, it's also less expensive — only $.05/minute" — the pricing detail the launch post buries; corroboration for the ai-file bundling story.
- ChatGPT Pro 20x purchases paused remotely (Tibor Blaho, 12:39 PM ET): "This plan is temporarily unavailable for new purchases. Existing subscriptions are unaffected." — a capacity/demand datapoint: the top tier is supply-constrained enough to throttle new buyers. Unverified motive (demand surge vs. staged rollout) — logged as a signal, not a fact.
- Perplexity API MCP server now supports OAuth (11:08 PM ET Sep 9):
https://api.perplexity.ai/mcpwith org sign-in — appends the perplexity-index thread: the Search API just got the lowest-friction integration path yet (MCP + OAuth = any agent harness can add it as a connector). The "index as agent infrastructure" thesis keeps executing. - Fable 5.1 Build Days (6:13 PM ET): Anthropic-community buildathons in cities worldwide Sep 11–25 — ecosystem-cadence note, the counter-image to OpenAI's Sep 16 GPT-6 community event.
Dropped: 49 filtered tweets (prompt-sharing, product chatter, engagement bait); ChatGPT/RT dupes of the ai-file stories (Financial Services, Data agent, GPT-Live-1) — covered there.