Models & releasesGoogle / Introducing Gemini 4 Neon

Gemini 4 Neon: check context, thinking and cost by channel

Google launched Gemini 4 Neon on November 18 for complex coding and agent workflows. Its API, Antigravity and Copilot availability differ, while migration requires attention to thinking defaults, shared context, caching and the cost of completed work.

A moss-green coding orc examines a glowing three-branch neon-glass sculpture beside a copper gauge, an empty checklist and lavender test blocks.
Conceptual illustration.

A new option for demanding, frequent developer work

Google introduced Gemini 4 Neon on November 18, positioning it between Hydrogen and Carbon for complex coding, analysis and agent workflows. Google claims faster completion and lower task costs than larger alternatives, alongside fewer unsupported claims during internal coding evaluations. Those are vendor findings, not measurements of a reader’s application.

The useful adoption question is specific: can Neon finish an existing repository task with an acceptable patch, a trustworthy completion report and a reasonable total cost? Keep those outcomes separate when comparing it with a model already in use.

Choose the channel before relying on its limits

The API model is gemini-4-neon. Its documented unified window is 1,048,576 tokens, shared across input, thinking and output; synchronous output defaults to a maximum of 65,536 tokens. Antigravity separately documents up to 512K context. Its Neon access includes paid Pro, eligible premium-device trials, Ultra and enterprise, with Free and Plus excluded.

GitHub lists Neon for Copilot Pro+, Max, Business and Enterprise, with gradual rollout and administrator policies controlling access. A model appearing in one interface does not establish access in another. Record the client, plan and observed model selection before comparing results, and check that the relevant workspace tools are available.

Review inherited API settings and conversation state

Google’s migration guide says another Gemini 4 model can switch to Neon without changing the request schema, but Neon defaults to high thinking and also supports low, medium and xhigh. For Gemini 3 migrations, minimal is rejected, explicit caching is unsupported, and the guide instructs developers to remove sampling parameters. In-range sampling values are ignored; out-of-range values fail validation.

Inspect what the wrapper actually sends. Preserve previous model turns and thought signatures when managing generateContent history, or use previous_interaction_id with the Interactions API. The guide also says max_output_tokens covers thinking and output together. An inherited small cap can therefore truncate the answer. Use a bounded cap suited to the task and inspect completion status.

Compare reasoning settings without changing the task

Antigravity labels Neon’s four levels Low, Medium, High and Extra; High is the default, and Extra corresponds to xhigh in its CLI. Its Orchestrator option plans and reviews with Neon while Hydrogen implements. Comparing that combination with standalone Neon also changes the execution setup.

For a useful trial, hold the repository revision, task instructions, tools and acceptance criteria fixed. Include a small repair, a multi-file change and a case with a missing prerequisite. Review the diff and actual test evidence, including whether the agent accurately reports work it could not finish. Repeat uncertain outcomes before drawing a reliability conclusion.

Calculate accepted-task cost and keep limitations visible

On November 19, Google’s Standard API table lists $3 per million input tokens, $0.15 for cached input and $15 for output, including thinking. Neon uses implicit caching only, so budget against observed cache hits rather than assuming repeated context always receives the discount. Grounding and other service tiers have separate pricing.

Count failed attempts, retries, tool charges and review time alongside successful runs. Lower token use alone does not establish lower cost per accepted change. The model card still identifies hallucinations as a limitation and gives a March 2026 knowledge cutoff. For current dependencies or API behavior, verify the generated references against the relevant documentation.

Keep six checks in the evaluation record

Codorc proposes this small comparison before expanding usage. No model requests, integration tests or benchmarks were executed for this article:

  • Record channel, client version, plan, model ID and permissions. Confirm actual access rather than inferring it from the launch announcement.
  • Capture the outgoing request, thinking level, output cap and context size. Review inherited caching and sampling fields.
  • Keep tasks and repository revisions fixed. Compare standalone and orchestrated workflows separately.
  • Inspect accepted patches and test output. Include a missing-file or unavailable-tool case and check the completion report.
  • Record actual input, cached input, output, elapsed time and total charges for every attempt, including unsuccessful ones.
  • Retain failure traces and the previous configuration. Set acceptance and stop conditions before directing more work to Neon.

Event dates refer to the original announcement. Publication dates show when this article first appeared on Codorc.

Continue reading

All stories
A moss-green coding orc holds a lavender bridge tile above two copper tool-carts stopped at gaps in ceramic tracks, beside a blank checklist and copper weights.
Models & releasesOpenAI / GPT-6.2 launch

GPT-6.2: review request history before changing models

OpenAI released GPT-6.2 Astra and Sol on November 17. Beyond their claimed coding improvements, API adopters need to review reasoning settings, conversation-prefix changes, instruction placement and the cost of their actual workload.

A moss-green coding orc sorts a copper envelope into an open dispatch slot beside a closed lavender shutter and a tray of blank receipts held by copper paperweights.
Models & releasesHugging Face

Hugging Face’s OpenRouter switch: check routing and billing

Hugging Face now routes language-model Inference Providers traffic through OpenRouter. The November 16 change also removes custom provider keys for those models, making provider selection, the billed account and failure behavior worth reviewing before a production rollout.

A moss-green robot turns a copper dial above four lavender lenses focused on a puzzle, beside a single glass chamber of ceramic token blocks.
Models & releasesGoogle

Gemini 4 Hydrogen: check thinking levels and migration limits

Google announced Gemini 4 Hydrogen on November 12 for coding and agentic workflows. Before switching an existing integration, check its thinking settings, shared token budget and request compatibility against your own tasks.