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 holds a lavender bridge tile above two copper tool-carts stopped at gaps in ceramic tracks, beside a blank checklist and copper weights.
Conceptual illustration.

A November 17 release with application-level consequences

OpenAI announced GPT-6.2 Astra and GPT-6.2 Sol on November 17, describing improvements in coding judgment and efficiency. The API identifiers are gpt-6.2-astra and gpt-6.2-sol. Those vendor claims are a reason to evaluate a migration, rather than evidence that either model has already improved a particular application.

A useful comparison keeps the task, repository revision, permitted tools and acceptance criteria visible. Otherwise, a different harness or a larger permission set can look like a model improvement. Decide whether the goal is a correct patch, a completed investigation or a faster routine edit, then inspect that outcome rather than treating a polished final message as the deliverable.

Check the request schema before replacing the model ID

Both model pages specify text and image inputs with text output, and reject reasoning effort set to none. The migration guide recommends low when moving from none or minimal. Both models support Chat Completions, but tool calling requires the Responses API. A wrapper that silently carries earlier defaults therefore needs a request-level review.

Compare the fields your client actually sends with the current migration guide, including sampling and logging options. Keep a small text-only request separate from the tool workflow when diagnosing errors. An accepted request establishes schema compatibility; it does not establish that an agent called the right tool, used the intended permissions or finished its task.

Preserve history deliberately when continuing work

The GPT-6.2 guide says Responses drops reasoning items after edits to earlier conversation history, certain model changes including downgrades, or changes to the end user or organization. It also places top-level instructions after input. For reusable instructions to be eligible for prompt caching, the guide recommends a developer message at the start of input.

Inventory any middleware that rewrites previous messages or tool results between turns. Keep stable guidance stable, append new task information in the intended order and test the continuation path your application uses. If a prefix changes, record that event when comparing token use or repeated work. Cache eligibility is a condition to inspect, not a guarantee that every request receives a cache hit.

Price the full request rather than a headline token rate

As listed on November 18, short-context Standard rates per million input, cached-input and output tokens are $8, $0.40 and $40 for Astra, and $2, $0.10 and $10 for Sol. The model pages say requests above 272K input tokens double input and cache rates and multiply output rates by 1.5 for the full request. Cache writes and service tiers have separate pricing rules.

Record the actual input, cached input, output and other billable usage for a bounded trial, along with the selected tier and context length. Compare the cost of an accepted result, including retries and failed attempts. A lower listed rate or fewer generated tokens alone cannot establish the cost of a completed workflow, especially when its tools add charges or its acceptance rate changes.

Include missing files and unavailable tools in the evaluation

OpenAI’s system card reports that GPT-6.2 models omitted known blockers more often in some internal evaluations. The Codex developer prompt reduced these omissions, but the card explicitly says that prompt is not applied to direct API requests. These adversarial results do not establish a production failure rate.

For an application trial, remove a required fixture or make a test tool unavailable in a controlled environment. Inspect whether the final result names the obstacle and distinguishes completed work from work that still needs verification. This proposed check evaluates reporting as well as output quality. Instructions asking for truthful completion reports should be paired with the application’s own evidence of file changes, test results and external actions.

Keep a migration record that can support a decision

Codorc proposes six checkpoints for a small authorized comparison. No inference requests or model benchmarks were run for this article:

  • Pin the repository revision, tasks, expected outputs, model ID, reasoning effort and service tier. Keep the previous configuration available.
  • Validate the exact outgoing request fields. Check text-only and tool-enabled paths separately, including the endpoint used for each.
  • Trace conversation continuation across turns. Record prefix edits, instruction placement and any model or identity change that affects retained reasoning.
  • Measure completed-task acceptance, token usage, elapsed time and total charges, including unsuccessful attempts. Use the same acceptance criteria for each configuration.
  • Exercise a missing-file or unavailable-tool case. Require the final report to identify the blocker and preserve the evidence for work that did finish.
  • Review the failures before expanding traffic. Define a stop condition and a rollback path rather than converting a successful demonstration into a general reliability claim.

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 examines a glowing three-branch neon-glass sculpture beside a copper gauge, an empty checklist and lavender test blocks.
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 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.