GPT-6-Astra Goes Live: Codex Model Picker and Amazon Bedrock Catalogs

GPT-6-Astra is OpenAI's next-generation flagship model, and as of 2026-09-09 it is selectable: Codex CLI 0.154.0 lists it first in the model picker, Amazon Bedrock catalogs now carry openai.gpt-6-astra, and a bundled official migration guide confirms that the Responses API accepts model: "gpt-6-astra" directly. Only six days passed between this and the model's first appearance in the SDK type layer (openai-python v3.8.0 / openai-node v7.10.0 on 2026-09-03, covered in gpt-6-astra appears in the OpenAI SDK: the new ChatModel ID and the Safety Alerts API).

Every fact in this article comes from GitHub artifacts loadable on 2026-09-15: release notes, PR diffs, and bundled files at the rust-v0.154.0 tag of the openai/codex repository. The official OpenAI docs domains (openai.com / developers.openai.com) still returned 403 to this site that day, so any status touching those pages is date-anchored, and where no pricing figure exists, we say so.

1. Overview: a six-day timeline from type layer to selectable

The CLI release notes close the chain in one sentence, and every link in that chain can be pointed to verbatim inside the repository — which is exactly how this article is written.

2. The picker and Bedrock: what each PR did

2.1 Model picker: visibility=list, ranked first

PR #42879 is a small change with a strong signal: it set GPT-6-Astra's bundled model visibility to list so it appears in the interactive model picker — in first position. The models.json file at the rust-v0.154.0 tag cross-checks this: gpt-6-astra carries priority: 1, while the GPT-5.6 trio sits at priorities 6, 7, and 8.

The same entry sets minimal_client_version: 0.153.0 — the minimum client version the entry supports.

2.2 Amazon Bedrock: openai.gpt-6-astra with two regional variants

PR #42619 (merged the same day the SDK type layer appeared) added openai.gpt-6-astra to the Amazon Bedrock model catalog, with global and US cross-region variants in the Bedrock Runtime catalog, applying Bedrock-specific capability identifiers and context-window overrides. OpenAI model families have landed on Bedrock before — GPT-5.4 and GPT-5.5 took the same route in June 2026.

3. What the bundled metadata says

Key fields of the gpt-6-astra entry in models.json (rust-v0.154.0 tag, checked 2026-09-15):

3.1 Six reasoning levels, with ultra meaning task delegation

The supported_reasoning_levels for gpt-6-astra, each with its official description:

Two things stand out: the none effort is unsupported (stated explicitly in the guide's Limitations section), and ultra's description adds "automatic task delegation" beyond max — reasoning plus automatic task hand-off. The Codex system prompt base also changed in this version to "You are Codex, an agent based on GPT-6".

3.2 The 272K/872K figures are Codex operating windows, not an API model card

models.json is Codex's own runtime configuration, not an API model card: the GPT-5.6 trio shows the same 272K/872K windows in this file while their established API setting is a 1.05M context. This article therefore does not present those numbers as gpt-6-astra's "context specification" — they describe how large a session runs in Codex, not what the API allows.

4. The bundled official migration guide, distilled

PR #42931 replaced the bundled GPT-5.6 Sol migration guide in the OpenAI Docs skill with a GPT-6 Astra edition (upgrading-to-gpt-6-astra.md). The file is a snapshot of the official guide whose header defers to the live document (the live address is the latest-model/gpt-6-astra page on developers.openai.com, which returned 403 to this site that day). Four parts of the snapshot matter most:

4.1 Four new capabilities

• Async tool calling: set async: true on a function or custom tool. Your application still executes the tool and manages pending work; the model keeps reasoning, calls other tools, or answers independent parts of the request while it waits. Return the result with the original call_id when ready.
• Mid-turn steering: send additional user instructions while the model is working — a correction or a change in requirements — over a WebSocket connection. The Responses API preserves completed work and includes the update in a continuation.
• configuration_update input items: raise or lower reasoning.effort mid-conversation without rewriting the original prompt prefix, preserving cache hits; the new effort applies until another configuration_update overrides it.
• Misalignment monitoring: systems asynchronously monitor for misalignment and trigger alerts when necessary — the same thread as the Safety Alerts API that appeared in the SDKs in early September.

4.2 Limitations and the migration cleanup checklist

4.3 Role mapping: Sol moves to Astra; Terra and Luna stay

The guide maps migrations by role: GPT-5.6 Sol or an earlier flagship goes to gpt-6-astra ("Astra is the flagship-equivalent tier"); balanced workloads stay on Terra; speed- and cost-sensitive routes stay on Luna ("retain Terra for balanced work and Luna as the primary faster or cheaper model"). Pickers and registries should append Astra while keeping Terra and Luna entries unless the user explicitly asks for replacement.

5. Why pricing still reads "unverified"

The snapshot keeps exactly one qualitative sentence: Astra achieves stronger results with substantially fewer output tokens, delivering a lower estimated API cost per task than earlier models despite its higher per-token pricing — no numbers anywhere. As of 2026-09-15, openai.com and developers.openai.com both return 403 to this site, the pricing page cannot be checked, and no GitHub-side artifact contains a price. So this article states plainly: no pricing figure has appeared in an official channel; budget assessments should wait for the official pricing page.

6. How to verify this yourself

From any environment that can reach GitHub, four commands reproduce the core facts:

# Fact 1: the Astra entry in the CLI 0.154.0 release notes
curl -s https://api.github.com/repos/openai/codex/releases/tags/rust-v0.154.0 | grep -i astra

# Fact 2: window, priority, and reasoning levels in models.json
curl -s https://raw.githubusercontent.com/openai/codex/rust-v0.154.0/codex-rs/models-manager/models.json | python3 -m json.tool | grep -A 30 '"slug": "gpt-6-astra"'

# Fact 3: the full bundled migration guide
curl -s https://raw.githubusercontent.com/openai/codex/rust-v0.154.0/codex-rs/skills/src/assets/samples/openai-docs/references/upgrading-to-gpt-6-astra.md | head -40

# Fact 4: the two PRs behind the picker and Bedrock entries
curl -s https://api.github.com/repos/openai/codex/pulls/42619 | python3 -c "import json,sys; print(json.load(sys.stdin)['title'])"

One aside: models.json also contains two visibility: hide entries, gpt-daybreak-blue-latest / gpt-daybreak-red-latest, matching the Daybreak tiering OpenAI announced on 2026-08-07 (Blue for general defensive security work, Red for separately approved specialized models) — don't be surprised when the command above surfaces them.

7. Common misreadings and fixes

• Treating Codex windows as API specs: 272K/872K are Codex-side operating windows; the GPT-5.6 trio shows the same values in the same file. API-side specs live on the official model card (403 on the day of writing, unchecked).
• Passing ultra to non-Codex calls: ultra currently exists only in Codex bundled metadata; sending undocumented effort values to the Responses API carries no behavioral guarantee.
• Swapping the model name with sampling parameters attached: temperature, top_p, and top_logprobs are unsupported on gpt-6-astra; clean them per the 4.2 checklist first.
• Using tool calling on Chat Completions: the official guide says tool calling requires the Responses API — migrate the API shape before swapping the model.
• Enabling fast mode under EU data residency: officially unavailable; don't assume a latency tier exists in compliance-sensitive deployments.
• Switching everything at once: the guide itself recommends role-based mapping, appending rather than replacing, and its five-step validation matrix (old model and prompt, new model preserving effort, one effort lower, minimal fix, isolated optional features).

8. Next steps

• gpt-6-astra appears in the OpenAI SDK: the new ChatModel ID and the Safety Alerts API: the previous chapter — the type-layer debut and the Safety Alerts API.
• Codex CLI 0.154.0 and SDK 0.154.0: Worktrees, ExternalMessage, and the ultra Reasoning Effort: the rest of the same release.
• OpenAI Models Release Notes (2026, Living Document): the full model timeline, with the Astra entry synced.
• The complete guide to GPT models (2026-07): GPT-5.6 Sol, Terra, Luna: current selection guidance with Terra and Luna roles intact.