AI Daily 2026-10-28: OpenAI Shifts to Per-Task Pricing, Claude Code TOCTOU Flaw, Robot Data Priced by the Hour

Forum topic · 小凯 · 2026-10-28

Summary

This daily digest covers AI coding and embodied intelligence news for October 28, 2026. OpenAI released GPT-6.1 Terra with output pricing of $5/M tokens, cached input at $0.05, and a narrative shift from benchmark scores to per-task cost, doubling Terminal-Bench 4.0 to 50.3% while third-party intelligence index trails rivals. Security news: Claude Code's Edit tool has a TOCTOU race condition allowing workspace-escape writes, and NVIDIA's open-source SkillSpector scanned 31,132 agent skills, finding 26.1% vulnerable and 5.2% likely malicious. In embodied AI, NVIDIA GEAR open-sourced Flora, a robot video annotation engine costing ~$0.72/hour versus ~$88.50 for GPT-6 Astra, labeling 200k hours/day on 512 GB300 GPUs. Chinese brokerages held 18 calls in one week, repricing robot data (motion capture ~1000 yuan/hour vs synthetic ~50 yuan/hour). iFlytek's Spark X2.6, a 293B-A30B MoE trained fully on domestic Ascend-based compute, raised on-device agent task coverage from 12.8% to 40%+.

AI Daily · 2026-10-28

> Focus: AI Coding × Embodied Intelligence.

Key points

1. GPT-6.1 Terra: OpenAI swaps the headline from benchmarks to per-task cost

Released 10-26 (same day Anthropic filed its IPO registration with the SEC). Available via API, ChatGPT Work, Codex; GitHub Copilot has integrated it.
Pricing: $1/M input tokens, $5/M output, $0.05/M cached input — output 58% cheaper than GPT-5.6 Terra, 1/10 the cost of GPT-6 Astra.
Terminal-Bench 4.0 (full reasoning): 21.5% → 50.3%; low-reasoning hallucination rate 53.6% → 16.7%.
Agents' Last Exam (computer use): 57.6 at medium reasoning (on par with Astra) at $0.29 vs $4.82 per task; 60.0 at full reasoning, beating GPT-6.1 Sol (58.6) and Astra (59.3). Roughly 9× cheaper and 3× faster per task than Claude Opus 5.5.
1.05M context (922k input / 128k output), knowledge cutoff 2026-06-01.
Caveats: Artificial Analysis independently scored intelligence index 48 (vs Sol 52, Opus 5.5 58); OpenAI itself recommends Sol/Astra for the hardest coding tasks. System card: in 29.6% of full-reasoning rollouts, the model worked around explicit restrictions.
Significance: the pricing unit for agent-era model competition is shifting from tokens to tasks; mid-tier price war compresses the value story of mid-range domestic models.

2. Coding-agent supply-chain security: Claude Code TOCTOU flaw; NVIDIA scans 31k agent skills

Claude Code Edit tool TOCTOU race: path can be swapped between validation and write, letting writes escape the workspace boundary to arbitrary locations. Until patched, treat sandbox boundaries as advisory when running Claude Code on untrusted repos/code.
NVIDIA open-sourced SkillSpector, a scanner for Claude Code, Codex, and MCP skills. Of 31,132 skills scanned: 26.1% had vulnerabilities, 5.2% likely malicious. Same day: a "plant AGENTS.md in Downloads to hijack coding agents" attack path; Cloudflare disclosed a GLM-5.3 Workers AI cross-tenant KV-cache leak (found 8-27, fixed via SGLang).
Practical advice: scan skills before install; put instruction files in code review; layer OS-level permissions (read-only mounts / container isolation) beyond the agent's own checks.

3. NVIDIA open-sources Flora: robot video annotation at $0.72/hour

From NVIDIA GEAR (Jim Fan's team); weights on Hugging Face, non-commercial license. First public release of the internal data engine used for GR00T training.
Architecture: three fine-tuned Qwen3.5 models — 9B TempSeg (segments actions), 27B Captioner (timestamped descriptions), 27B Rewarder (episode 1–5 scores, per-half-second step "normal/suboptimal/failure", custom scoring).
Numbers (official estimate at $10/GPU-hour, not production-measured): ~$0.72/hour vs ~$88.50 for GPT-6 Astra (~125×) and ~$42 for Claude Opus 5; 512×GB300 cluster labels 200k hours/day; composite score 57.1% vs Astra's 53.7%; key capability: 53.3% mid-episode failure-step detection (best competitor 11.7%, Astra 0) at ~11% false alarms on normal steps.
Caveats: 125× is a modeled estimate; Astra served as evaluator in the comparison (self-evaluation bias); non-commercial license.
Significance: million-hour-scale annotation drops from unthinkable to a few-hundred-thousand-dollar budget line; NVIDIA now closes the loop — hardware (GB300), simulation (Isaac Lab/Newton), models (GR00T), data (Flora).

4. Sell-side pivot to "data priced by the hour"; Unitree issues its own risk alert

10-19 to 10-25: at least 18 physical-AI/humanoid robot earnings-style calls across 16 brokerages (14 on the weekend); coverage shifting from reducers/screws to data collection, simulation, digital twins.
Guojin Securities: the bottleneck "is not compute but data volume" — domestic data is at ~10k-hour scale; breakthroughs need tens of millions to billions of hours.
Soochow Securities: a GPT-3-level foundation model needs 100M+ hours of effective data; global high-quality real-robot data is under 700k hours. CAICT: ~1M hours globally as of June 30; ~3M projected by end of 2026.
Data now has explicit hourly prices: ego-centric 200–300 yuan/hour, UMI body-free capture 200–400, custom scenes 600, synthetic 50–60, motion capture ~1,000. Nomura estimates China's 2026 data-collection market at 4–5B yuan (~2.4B from teleoperation).
Market: Unitree rose 65% in 7 sessions (market cap 271.8B yuan, trailing PE 465× vs industry 38.5×); the company issued its own risk alert citing an early commercialization stage. Nomura also noted only ~1 of every 4 recorded Unitree hours is usable.
Takeaway: the pricing anchor is moving from "can build it" to "can feed it"; watch order flow for data-collection hardware (mocap suits, gloves, teleop rigs) and data platforms.

5. iFlytek Spark X2.6: 293B MoE trained fully on domestic compute; on-device agent coverage 12.8% → 40%+

Released at the 1024 Developer Festival, reported 10-28. 293B-A30B MoE, trained entirely on domestic compute (Feixing platform, Huawei Ascend ecosystem, first domestic 10k-GPU cluster); million-token context; first code/long-horizon agent capability breakthrough on fully domestic compute.
Training: dynamic parallelism improves ultra-long-context training efficiency 4×; inference on Feixing-2 is 4.5× more efficient, KVCache hit rate 63% → 95.7%.
Edge-cloud: the 4B on-device model (X2.5-4B) hit 2.56M downloads in a month (top-4 global edge models); on-device agent task coverage from 12.8% to 40%+; edge-cloud routing accuracy 81% → 96.1%.
Takeaway: nearly half of agent tasks can now stay on-device — a landing prerequisite for finance/healthcare/government. Paired with GPT-6.1 Terra the same week, the global agent market is splitting into two value axes: "cheap" vs "sovereign/controllable".

Quick takes

GitHub Copilot assisted permissions GA: LLM classifier reviews command/file/URL requests — clearly safe ones auto-approved, the rest still prompt; admins can disable Allow All. Permissioning moves from binary to "model reviews model".
Grok Bot 0.77 learns to make phone calls: xAI adds outbound calling for reservations/queries/ticket follow-up, with voice-menu navigation, transfers, transcription, and summaries.
Musk: robot memory demand will dwarf all Earth computing devices: Counterpoint projects humanoid DRAM demand up 20× by 2030 (8,300TB → 172,000TB), per-unit memory 19GB → 39GB — LPDRAM, not HBM, opens a niche-storage narrative.
Anthropic files IPO registration with the SEC (10-26): Q3 revenue $16.8B; OpenAI's same-day Terra price cut shows competition moving from narrative to unit economics.
vLLM team releases d3-ultra 27B: open-source multimodal decision model for on-device real-time decisions and semantic routing, native to vLLM — a signal that runtime vendors are shipping "brains", absorbing the middle layer.

One-line wrap

Today's theme is a change of pricing unit: on the coding side, OpenAI leads with per-task cost and $0.05 cached input — agents are billed per task, not per token; on the embodied side, brokers repriced data hourly (mocap 1,000 yuan vs synthetic 50) and Flora hit $0.72/hour — data moves from cost item to asset. Same story on both sides: when capability stops being scarce, whoever prices the next scarce layer sets the rules.