Codex made GPT-6.1 Sol the default: what teams should change

A coding agent's default model is a setting most people never look at. When it changes, spend, speed and behaviour change with it, for everyone who did not pin a model. On 7 October 2026, Codex release 0.161.0 made GPT-6.1 Sol the default model in its bundled catalogue and its Amazon Bedrock catalogue. The same release changed how OpenAI's Daybreak cyber programme is switched on. Neither change is dramatic on its own, but both are the kind that quietly shifts what a team's agents do and cost.

What changed in Codex 0.161.0

The release notes for rust-v0.161.0, published on 7 October 2026, list these changes relevant to teams running Codex at scale:

• New default model. GPT-6.1 Sol is now the default in the bundled and Amazon Bedrock catalogues.
• Bedrock features. Amazon Bedrock supports multi-agent V2 and Ultra reasoning on compatible models, and Bedrock Mantle also accepts AWS GovCloud regions.
• Daybreak is opt-in. It is enabled only with --enable cli_daybreak or features.cli_daybreak=true; the notes say daybreak=true alone "is insufficient". By default its controls and indicators are hidden, /daybreak is unavailable, and automatic Cyber routing is omitted, even for saved Daybreak threads. Opt-in routing also requires an eligible ChatGPT sign-in, the OpenAI provider and advertised model and programme support.
• Per-turn cyber programme choice. A Cyber access programme can be selected per turn with codex exec --cyber-access-program or the TypeScript SDK's cyberAccessProgram option, without changing the saved choice.
• Permissions fixes. Approved filesystem escalation can now grant broader write access while keeping denied reads and network restrictions, and background tasks keep the permissions of the turn that started them.
• MCP sign-in. People can sign in to MCP servers from a running session with /mcp login.

What GPT-6.1 Sol is, and what it costs

OpenAI introduced GPT-6.1 Sol as an upgrade to GPT-6 Sol that "nearly matches" GPT-6 Astra on agentic coding, computer use and professional work, at one-fifth of Astra's standard input and output prices. It is available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.1-sol. A fuller comparison with Astra and Luna list prices is in GPT-6.1 Sol: near-Astra coding performance at one-fifth the token price.

The model page gives the standard API prices and some details that matter for agents:

Rate, per 1M tokens	GPT-6.1 Sol
Input	$2.00
Cached input	$0.10
Cache writes	$2.50
Output	$10.00

• Prompts with more than 272,000 input tokens are charged at twice the input and cache rates and 1.5 times the output rate, for the whole request.
• Reasoning effort supports low, medium (the default), high, xhigh and max.
• Batch and Flex prices are 50% lower than Standard; Fast mode is twice Standard, and Ultrafast mode six times.

One detail has moved since launch. The announcement said an Ultrafast option would come "in the coming days". The model page now documents Ultrafast mode and its price. If you read the announcement alone, you would think it was not yet available; check the model page for the current position.

For teams moving from GPT-6 Sol, OpenAI says cached input is now 50% cheaper than GPT-6 Sol's. For teams moving from a different default, the comparison is less simple: a model that uses more reasoning or more tool calls per task can cost more overall at the same per-token price.

Analysis: why a default change deserves attention

Silent upgrades change three things at once. Cost per task, latency and behaviour all shift when the model changes. A team that budgeted for one model, tuned prompts for its habits, or measured its pass rate on internal tasks is now running something different, without anyone deciding to.

Long prompts carry a price step. The 272,000-token threshold matters for agents working in large repositories. A session that pulls in many files can cross it, and then the whole request is charged at the higher rate. That is easy to miss if cost is only reviewed as a monthly total.

Daybreak becoming opt-in is the safer default. For most organisations, having cyber-specific routing appear only when someone deliberately enables it is the right posture. It also means anyone who relied on Daybreak before 0.161.0 now needs an explicit flag, and that the saved preference alone no longer turns routing on. Treat enabling it as a decision with an owner, not a per-developer toggle.

Bedrock users get a different set of changes. Teams running Codex against Bedrock should read the Bedrock lines carefully. Multi-agent V2 and Ultra reasoning apply only "on compatible models", and GovCloud support applies to Bedrock Mantle. These are not general statements about every Bedrock model.

Model choice does not change the sandbox. The Codex sandbox documentation describes the sandbox and approval policy as separate controls: the sandbox sets technical boundaries, and the approval policy decides when the agent stops and asks. Neither depends on which model is answering. A more capable default model is a reason to check those settings, not to relax them. The article on sandboxing coding agents covers what the built-in sandboxes do and don't cover.

What it means for engineering and security leaders

The general lesson is that any agent with a vendor-managed default model can change cost and behaviour on an ordinary update. For Codex specifically, this release is a good prompt to do three things.

First, decide where you want defaults and where you want pins. Interactive use by individual developers can reasonably follow the vendor's default. Continuous integration jobs, scheduled agents and anything with a fixed budget should name their model explicitly, so an upgrade is a choice you make and test.

Second, look at the opt-in programmes as policy, not preference. Daybreak and the per-turn Cyber access programme selection are capabilities with security implications. Write down who may enable them and where.

Third, watch spend in the weeks after an upgrade. A change in default model shows up first as a change in cost per session or per task.

Practical takeaways

• Pin the model explicitly in CI jobs, scheduled agents and shared configuration, rather than relying on the default.
• Re-run a small set of real tasks on GPT-6.1 Sol and compare cost per task, latency and pass rate with your previous default.
• Check whether any sessions exceed 272,000 input tokens, and how often; those requests are charged at a higher rate.
• If you used Daybreak before, confirm whether you still need it, and if so enable it deliberately with cli_daybreak.
• Decide who may use --cyber-access-program in automated runs.
• Review sandbox mode and approval policy after the upgrade; a stronger model is not a reason to widen them.

Where SpanDock fits

The SpanDock model proxy, in every plan, records the requested model and the model that answered for every call, with tokens and cost, and Codex can be pointed at it with the snippet SpanDock shows. A conversation keeps its model for 6 hours, so it doesn't change mid-task. Rules that swap one model for another are on the Premium plan and up, and SpanDock Mind and per-person budgets on Standard and up. To see Codex sessions on a dashboard in the first place, start with getting started.