Same Gemini 4 Neon, triple the coding score

Google's new midtier model solved 12.8% of a coding benchmark in a plain harness and 39.0% inside Google's own Antigravity. The model did not change. The harness did.
Gemini 4 Neon
- What it is: Google's midtier model, between the flagship Carbon and the cheap Hydrogen, and its fourth new Gemini 4 model in four weeks. About one million tokens of context in and out, text, images, video, audio and PDFs, Computer Use and code execution. No image or audio generation.
- Google's claim: most tasks about three times as fast as Carbon at about 60% lower cost.
- Independent number: Artificial Analysis Intelligence Index 61, third overall, behind Claude Fable 5.2 (65) and GPT-6.2 Astra (62).
- Price: $3 per million input tokens, $15 output, $0.15 cached input. GPT-6.2 Sol is $2 / $10, Claude Opus 5.5 is $4 / $20.
The harness did the work
From Google's model card, as The New Stack laid it out, on the SWE-sweep coding benchmark:
| Model | Plain mini-SWE harness | Maker's own harness |
| Gemini 4 Neon | 12.8% | 39.0% (Antigravity) |
| Claude Opus 5.5 | 3.6% | 10.1% (Claude Code) |
| GPT-6.2 Sol | 11.6% | 8.3% (Codex) |
Antigravity's harness was trained together with Gemini. Neon makes fewer unnecessary file reads and manages context better in long sessions, and Antigravity runs subagents in separate Git worktrees and checks their changes. Google tested Neon in only these two setups and did not separate which change helped, so there is no data yet on Neon in any other harness, including your own agent loop. A day earlier, the FrontierHarness eval found the same open model could cost up to five times more depending on the harness.
For builders: a benchmark score without its harness is half a number. Before you switch models, run your comparison inside the agent setup you actually ship, and record cost per finished task, not only the pass rate.
Banks put AI agent risk on you
Citibank changed its online user agreement: customers who give a third-party AI agent access to their account are responsible for what it does, and if the tool leaks credentials or is hacked, for what the hackers do. JPMorgan and Bank of America set similar rules last month, PNC changed its terms on Nov 18, and Citizens emailed customers the same warning. Citi's own AI agents carry no extra liability. If you build personal finance agents, read-only access is now the easy sell.
Nvidia after Huang
- Elon Musk at DealBook: ousting Jensen Huang is like kicking Michael Jordan off your team. The board says it is not bringing him back.
- Huang agreed to appear before the Senate Banking Committee, chair Tim Scott said; a closed-door interview with three House committees is set for Dec 1.
- The SEC opened a civil investigation into whether Nvidia misled investors about chip sales in Asia.
Security: unpatchable, and fast
- Satya Nadella said much of America's critical infrastructure is unpatchable: firmware, chillers, HVAC, all on the internet. With AI finding flaws faster, patches now have to go straight to production, which he says will create instability.
- OpenAI spent about $1.2 million in API costs to build working exploits for 948 Linux kernel flaws, out of 1,724 it reproduced. Humans reviewed findings before they went to the maintainers; OpenAI calls the cost an upper bound.
AI models as targeting tools
Vals AI tested models from eight providers on real and synthetic drone and satellite images. Most readily flagged civilians, ambulances, medics, journalists and surrendering soldiers as military targets, and no request was blocked by upstream safety filters. Refusals broke when the prompt called it a video game: GPT-6 Astra's target rate on drone images went from 6% to 92%. On real satellite images of a school in Minab, Iran, Claude Opus 5.5 marked the school as a military target in 70% of runs. OpenAI says it is updating its models; Anthropic said the behavior violates its rules and it is adding safeguards; Google pointed to its ban on autonomous weapons without human oversight.
Money
Anthropic priced its IPO at $83 a share, above the $70 to $78 range, and raised $81.3 billion, more than SpaceX's record in June. It trades on the Nasdaq as ANT.
Agents at work
- KPMG ran more than 100 AI agents through a full parallel audit of a FTSE 250 company. The official opinion came from the human audit; in places the agents went deeper, and in one area a human had to step in. KPMG plans up to 20 more.
- Hugging Face now routes language model requests through OpenRouter: `:fastest` and `:cheapest` suffixes, automatic fallback, no markup. Bring-your-own keys for language model providers stop working; image and video providers keep them.
The through-line
What wraps the model decides the result. A harness tripled a coding score, a video game framing unlocked targeting, and a bank's terms moved the risk to you. One thing to do today: run your model comparison again inside the harness you ship.
Sources covered
1 story · 148 sources (79 news, 63 company, 6 voices)
- llm-gemini 0.38🎙️Simon Willison↗🏢AWS AI/ML↗🏢AWS AI/ML↗🏢Snowflake↗🏢Databricks↗🏢AWS AI/ML↗📰The New Stack↗📰InfoQ↗📰SiliconANGLE↗🏢AWS AI/ML↗🏢Hugging Face↗📰The New Stack↗📰TechCrunch↗📰CNBC↗📰TechCrunch↗📰TechCrunch↗🏢LangChain↗🏢LangChain↗🏢LangChain↗📰SiliconANGLE↗📰SiliconANGLE↗🏢Hugging Face↗📰SiliconANGLE↗📰TechCrunch↗📰Hacker News↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗🏢MongoDB↗🎙️Simon Willison↗🏢Google Cloud AI↗🏢Meta AI↗🏢Meta AI↗📰Hacker News↗🏢LangChain↗🏢Starburst↗🏢Snowflake↗🎙️Aurimas Griciunas↗🏢Starburst↗📰The Decoder↗🏢Starburst↗🏢Pinecone↗🏢Databricks↗🎙️Simon Willison↗📰BleepingComputer↗📰The New Stack↗📰The New Stack↗📰TechCrunch↗🏢Hugging Face↗🏢Confluent↗📰TechCrunch↗🎙️Simon Willison↗📰RIA Novosti↗📰TechCrunch↗📰Engadget↗📰The Verge↗📰BleepingComputer↗📰The Verge↗📰The Decoder↗📰Ars Technica↗📰SiliconANGLE↗📰SiliconANGLE↗📰The New Stack↗📰The Verge↗🏢Databricks↗🏢NVIDIA↗📰The Decoder↗📰The Verge↗🎙️Ben Lorica↗📰The New Stack↗📰Ars Technica↗📰TechCrunch↗📰The Decoder↗🏢MongoDB↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰SiliconANGLE↗📰TechCrunch↗📰SiliconANGLE↗📰The New Stack↗📰RIA Novosti↗📰The Decoder↗📰The New Stack↗📰Hacker News↗📰The Decoder↗📰The Verge↗🏢OpenAI↗📰SiliconANGLE↗📰MIT Technology Review↗📰The Decoder↗📰The Decoder↗📰The Decoder↗🏢ClickHouse↗📰The Decoder↗📰SiliconANGLE↗📰The Verge↗📰The New Stack↗📰SiliconANGLE↗📰TechCrunch↗📰The Hackers News↗📰Boca Raton Tribune↗📰The New Stack↗📰The Verge↗📰SiliconANGLE↗📰TechCrunch↗🏢Vercel AI↗🏢OpenAI↗🏢OpenAI↗🏢OpenAI↗📰The Verge↗📰InfoQ↗🏢OpenAI↗🏢Fivetran↗🏢Fivetran↗🏢Fivetran↗🏢Fivetran↗🏢Fivetran↗📰RIA Novosti↗📰The Decoder↗📰MIT Technology Review↗📰InfoQ↗🏢Confluent↗🏢Vercel AI↗🏢Vercel AI↗📰TorrentFreak↗🏢Databricks↗📰Hacker News↗🏢LangChain↗🏢Polars↗🏢Modal↗🏢Temporal↗🏢Temporal↗🏢Temporal↗🏢Temporal↗🏢Polars↗🏢Elastic↗🏢Elastic↗🏢PlanetScale↗🏢vLLM↗🏢Qdrant↗🏢MotherDuck↗🏢Vercel AI↗🏢Vercel AI↗🏢ClickHouse↗🏢Hugging Face↗🏢Hugging Face↗🏢OpenAI↗