codex 0.145: bedrock login, realtime audio, and sub-agent controls in one release

OpenAI released Codex v0.145.0 on July 21, 2026 — one of the larger single-version drops since the product launched. the headline additions are Amazon Bedrock as an authentication and endpoint target, audio inputs with streaming Realtime V3 conversations, paginated thread history, and a sub-agent orchestration layer with configurable models and concurrency. underneath, the /import command grew to cover Cursor and Claude Code migrations end-to-end.

amazon bedrock login

Codex can now authenticate against Amazon Bedrock directly, including custom endpoint and authentication configurations. GPT-5.6 Sol is the default Bedrock model when using this path — the same Sol that became the default in standard Codex sessions after the GPT-5.6 family launched on July 9.

the practical impact: enterprise teams running Codex inside an AWS environment can keep inference on their existing Bedrock contracts, data residency controls, and IAM authentication rather than routing through OpenAI's API. for organizations that already have Bedrock integrated into their compliance and cost management stack, this removes a significant friction point.

Bedrock support is marked experimental in v0.145.0. the feature exists, but production workloads should validate behavior against their specific endpoint and model configuration before relying on it.

audio inputs and realtime v3

Codex now accepts audio inputs, including common local audio formats. separately, the release introduces streaming Realtime V3 conversations — a persistent connection mode where Codex's responses stream in real time rather than completing before rendering.

this builds on the Voice integration added to the desktop app in the previous Codex update, but goes further: the Realtime V3 path is a lower-latency channel that keeps the conversation alive rather than treating each voice exchange as a discrete request. for developers using voice as a control surface for agent runs, the difference is between walkie-talkie and phone call.

paginated thread history

Codex sessions have historically been ephemeral by default — there was no built-in way to search past conversations or resume from where you left off across sessions. v0.145.0 adds experimental paginated thread history with efficient resume, search, and persisted names.

the "efficient resume" part matters specifically: rather than re-sending the full prior context on every session open, the history is fetched incrementally. for long-running codebases where a thread has hundreds of exchanges, the difference in startup latency is meaningful. sub-agent records are also included in thread history, so you can trace what sub-agents did in a prior session the same way you can trace what the primary agent did.

sub-agent orchestration controls

the opt-in multi-agent V2 experience that shipped earlier in July is now stabilized in v0.145.0 with a set of explicit configuration controls: per-sub-agent model selection, reasoning level, and concurrency limits. restored roles let you define what scope each sub-agent is allowed to act in.

configurable concurrency is the key addition for teams with rate limits or cost constraints. previously, multi-agent Codex would spawn sub-agents up to the global concurrency cap without any project-level throttle. now you can cap sub-agents per session and specify which model runs at which reasoning level — running a cheap fast model for parallelizable file reads and a frontier model for the synthesis step, for example.

/import migration from cursor and claude code

the /import command has been substantially expanded. it can now migrate full Cursor setups — rules, MCP servers, plugins, and project-scoped memories — as well as Claude Code configurations including memory files, MCP connections, and session history.

this is the first time OpenAI has built a tool explicitly targeting migration from competitors' products. the framing is practical (developer experience) rather than competitive (lock-in), but the effect is the same: the cost of switching from Cursor or Claude Code to Codex just dropped substantially. whether it drops enough to matter depends on how much Cursor-specific behavior the migration faithfully replicates, which will take the community some time to fully evaluate.

windows and performance fixes

v0.145.0 also includes improved Windows reliability — Codex has had rough edges on Windows since launch, and this release addresses several of the higher-frequency failure modes. terminal performance improvements are included as well, though OpenAI hasn't published specific latency numbers.

taken together, the 0.145.0 release reads like a maturity push: history, configurable orchestration, and enterprise auth address the gaps that kept Codex in the "impressive demo, hard to operationalize" category for some teams. Bedrock support in particular suggests OpenAI is prioritizing enterprise deployment paths over the next few quarters.