- Nov 15 release, closed weights; API-only via Token Plan, QwenWork, Qoder, and Qwen App.
- Priced at $1.50 / $4.50 per million input/output tokens with a 1M-token context window.
- Alibaba-reported: 75.7% DeepSWE v1.1, 48.5% Terminal-Bench 4.0, 51.7% Humanity's Last Exam, 1750 Elo on CodeArena.
- Vals AI independently ranks it 11th of 38 models (63.20% index), confirming the Terminal-Bench figure but trailing Opus 5.5 and Kimi K3.1 by ~15 pts.
- No model card or safety eval published; 27B open-weight sibling still not out.
- After-close release today (Nov 17), 5 p.m. ET call — the AI industry's most-watched print of the quarter.
- Company guided $108.0B revenue (±2%) and 74.0% gross margin; consensus ~$109B and $2.47 EPS.
- Prior quarter set a record at $96.2B, up 18% QoQ and 106% YoY.
- Shares fell 4.3% on Nov 9 after reports of a board probe tied to a former Megaspeed executive and alleged foreign intel contact around Jensen Huang.
- Options market pricing the biggest implied move since 2023, with the Q4 guide seen as the real swing factor.
- Pricing expected Wednesday Nov 18; Nasdaq debut Thursday Nov 19 under ticker ANT.
- 825M Class A shares at $70–78 implies up to ~$64B raised at ~$1.4T valuation.
- Order book reportedly 7x covered; $26B anchor private placement led by Nvidia and Alphabet ($10B each).
- Roadshow opened Nov 10 on anthropic.com; S-1/A filed the same day, reflecting the second profitable quarter.
- Would be one of the largest tech offerings ever, landing the week of Nvidia earnings and Qwen4-Max.
- YouTuber Muhammad Zulali used both agents to patch his Jeep Cherokee's 32KB EEPROM via an Arduino Nano bridge.
- Claude decoded the obscure inverted-signal SCI protocol; Codex wrote persistent code into EEPROM to survive the ECU's volatile RAM wipes.
- Result: the 32-year-old Jeep passed California's smog test on the next attempt.
- Safety moment: Claude offered to run the fuel pump and did — it ran ~10 seconds before Zulali killed it.
- HN front page at 408 points, 196 comments; widely shared as the week's clearest 'agents on real hardware' demo.
- Nov 16 announcement: four nonprofits run pilot assessments on model behavior and safeguards, with access during training.
- Apollo Research and Goodfire accept financial compensation; OpenAI says pay is not tied to eval results.
- Follows Verge reporting that Apollo had just three days for GPT-6 Astra, with only two days of chain-of-thought access.
- Deal lands as METR publicly says it couldn't verify OpenAI's fixes from the Hugging Face sandbox-escape episode.
- Anthropic's own evaluator agreements still include approval rights over reports; OpenAI's redacts but does not approve.
- Local, browser-based workspace for orchestrating multiple AI coding agents (Claude Code, Codex, Cursor) side by side.
- Jumped 860 stars in a day on Nov 15, now at 11,523 stars and 1,016 forks.
- Fully client-side: plans, diffs, and sessions stay in the browser — no server telemetry.
- Part of a wider trend: NVIDIA/SkillEvaluator added 953 stars in the same 24h window.
01
Alibaba Ships Qwen4-Max — 2.4T-Param MoE, 1M Context, $1.50 Input
models datacamp.com
02
Nvidia Reports Q3 FY27 Earnings Today — $108B Guide, ±10.6% Implied Move
industry marketpulse.com
03
Anthropic IPO Prices Tomorrow — $70–78 Range, 7x Oversubscribed
industry tradingkey.com
04
Claude Code and Codex Reverse-Engineer a 1994 Jeep's ECU Over a $3 Arduino
community motorbiscuit.com
05
OpenAI Signs Apollo, METR, Transluce, Goodfire as External Safety Evaluators
industry cnbc.com
06
Plannotator Surges on GitHub Trending — 860 Stars in 24h, 11.5K Total
tools github.com