Codex, Copilot, and Claude Code Now Choose Your Model and Provider

Today is about who chooses your model and provider. Codex put GPT-6.1 Sol at the top of its bundled catalog the day OpenAI released it, so a session that never named a model may start on a different one than last week, while GitHub added it to Copilot and Claude Code gained a way for administrators to say which API providers a machine may call.

Treat today as the day you name your model and your provider on purpose: pin gpt-6.1-sol in Codex or stay on the model you tested, decide whether Copilot’s new row should be on for your org, and if you try allowedProviders, put it in the managed settings file.

Top Stories

1. GPT-6.1 Sol arrives, and Codex makes it the default

Practical dev impact: OpenAI’s model page lists gpt-6.1-sol at $2 per million input tokens and $10 per million output, the same list prices as GPT-6 Sol. The change is caching: cached input drops to $0.10 per million, half of GPT-6 Sol’s $0.20, and cache writes cost $2.50. The catch is that a prompt over 272K input tokens is billed at twice the input and cache rates and 1.5 times the output rate for the whole request, so long-context agents should measure how often they cross that line. Reasoning effort accepts low, medium (the default), high, xhigh or max, not none or minimal, so check what your code sends. OpenAI lists it for Plus, Pro, Business, Enterprise and Edu plans in ChatGPT Work and Codex and in the API, but not yet in Chat.

OpenAI says GPT-6.1 Sol matches GPT-6 Astra on DeepSWE v1.1 at roughly one fifth of the cost and beats GPT-6 Sol’s best score by 6.4 percentage points, and that its factual-error rate at low effort falls from 11.4% to 7.7% on prompts it calls deliberately difficult and not representative. Those are vendor numbers, so run your own evaluation. GPT-6.1 Sol Ultrafast is promised only for the coming days.

In Codex, “default” is narrow. The v0.159.1 release (published 4:32 PM ET on 29 Sep) lists GPT-6.1 Sol as the default model in the bundled catalog, and on v0.159.2 it is the first entry by priority in codex debug models --bundled. That is a catalog ordering, not a config key. Release v0.159.2 (7:57 PM ET) holds the same content plus a backport that suppresses Windows console windows for background processes and sandboxed commands, so the Sol default does not need v0.159.2, which is just the version to try on Windows. The backport’s pull request says native Windows execution was not validated on that candidate and issue #48074 was still open this morning, so test before you call it fixed. To choose the model yourself, pass codex -m gpt-6.1-sol or set a top-level model = "gpt-6.1-sol" in ~/.codex/config.toml.

2. Copilot ships GPT-6.1 Sol to the paid plans above Pro

Practical dev impact: GitHub’s changelog, dated 29 Sep with no time of day, says GPT-6.1 Sol is generally available and rolling out gradually in Copilot for Pro+, Max, Business and Enterprise, and Pro is not on that list. It appears in the model picker across the IDEs, Copilot CLI, the coding agent, github.com and GitHub Mobile, billed at the provider’s list price under usage-based billing. GitHub says its early testing found noticeably fewer tokens and steps than earlier GPT-6 and GPT-5.6 models, which is GitHub’s own claim. Business and Enterprise admins manage access through the model policy in Copilot settings, and under default model enablement a new model is on unless the admin turned off the global default or disabled this model by name.

3. Claude Code v2.1.285 adds a provider allow list and a Desktop handoff

Practical dev impact: The changelog says v2.1.285 (published 3:27 PM ET on 29 Sep) added an allowedProviders managed setting that limits which API providers a machine may use. Anthropic’s docs do not describe it yet, so the details below come from the changelog and the settings schema shipped in the v2.1.285 package, and they may change. The schema says the setting is honored only from managed settings, meaning managed-settings.json, device management or server-managed settings. The value is an array of provider names such as anthropic, vertex and gateway. Unset allows every provider and unknown names are ignored, while an empty list or one with only unknown names allows none, so Claude Code refuses to start. A session on an unlisted provider is refused at startup, at login and on its next API call.

The catch is delivery. The schema says only a list in managed-settings.json or device management is enforcement on the machine, and a list set only in the admin console reaches just the sessions that fetch server-managed settings. It misses sessions on a cloud provider, another organization or a non-Anthropic ANTHROPIC_BASE_URL, sessions that authenticate only with apiKeyHelper or ANTHROPIC_AUTH_TOKEN, Pro or Max logins, and a first launch before the fetch lands. Versions older than v2.1.285 ignore the setting, so the schema suggests a minimum-version policy on a mixed fleet. If the OS denies read access to the managed file, v2.1.285 now warns and starts without that file’s policies.

The other headline item, claude --desktop, opens the Claude Desktop app on the current directory, or on a session with --continue or --resume <id>, and it is not in the docs either. In the v2.1.285 build it works on macOS and Windows x64 only, needs a Claude account sign-in, takes no prompt and cannot be combined with --bg, --worktree or --tmux, so keep it out of CI.

Practical Impact Analysis

OpenAI and GitHub changed what happens when nobody chooses, and Anthropic added a way to choose for a whole machine. A Codex session with no model set can start on GPT-6.1 Sol, a Copilot org on default enablement can find a new row switched on, and Claude Code has a provider list that counts as machine policy only in the managed file.

Name the Codex model in config so the next catalog change does not move you, decide on Copilot’s model policy before developers find the new row, and stage allowedProviders on a test machine first, because a list that omits a provider someone uses will stop them from starting Claude Code.

Tutorial

Check the Codex catalog default, decide whether to pin, and stage a provider list. This takes about fifteen minutes, needs bash, npm and jq, installs Codex into a temporary folder and writes one example file in the current directory.

1. Install Codex v0.159.2 into a temporary folder and print the first model in its bundled catalog, which should be gpt-6.1-sol. The command is a debug command and may change.
2. Decide whether to pin. A top-level model = "gpt-6.1-sol" line in ~/.codex/config.toml holds the model whatever the catalog default becomes, and you make that edit by hand.
3. Write the example managed settings file and check that it holds a non-empty array. Edit the provider names to match your approved backends.
4. On a test machine, copy the file to the managed path with admin rights: /etc/claude-code/managed-settings.json on Linux and WSL, /Library/Application Support/ClaudeCode/managed-settings.json on macOS, or C:\Program Files\ClaudeCode\managed-settings.json on Windows. Confirm v2.1.285 or newer with claude --version, then read the Setting sources line in /status.
5. Try a session on a provider you left off and expect a refusal, then confirm every developer’s real provider is listed before rollout.

#!/usr/bin/env bash
set -eu

# 1) Codex: install v0.159.2 in a temp folder, then read the first catalog entry
CODEX_DIR="$(mktemp -d)"
npm install --prefix "$CODEX_DIR" @openai/codex@0.159.2
"$CODEX_DIR/node_modules/.bin/codex" --version
"$CODEX_DIR/node_modules/.bin/codex" debug models --bundled \
  | jq -r '.models | sort_by(.priority) | .[0].slug'

# 2) Optional pin, done by hand: add this top-level line to ~/.codex/config.toml
#    model = "gpt-6.1-sol"

# 3) Claude Code: write an example allowedProviders file and check it
cat > managed-settings.example.json <<'JSON'

... click "Show full code" below to expand

#!/usr/bin/env bash
set -eu

# 1) Codex: install v0.159.2 in a temp folder, then read the first catalog entry
CODEX_DIR="$(mktemp -d)"
npm install --prefix "$CODEX_DIR" @openai/codex@0.159.2
"$CODEX_DIR/node_modules/.bin/codex" --version
"$CODEX_DIR/node_modules/.bin/codex" debug models --bundled \
  | jq -r '.models | sort_by(.priority) | .[0].slug'

# 2) Optional pin, done by hand: add this top-level line to ~/.codex/config.toml
#    model = "gpt-6.1-sol"

# 3) Claude Code: write an example allowedProviders file and check it
cat > managed-settings.example.json <<'JSON'
{
  "allowedProviders": ["anthropic", "vertex"]
}
JSON
jq -e '.allowedProviders | type == "array" and length > 0' managed-settings.example.json

# 4) On a test machine, copy the file to the managed path for your OS, then run
#    claude --version and open /status inside a session (both done by hand)

You are done when step 1 prints gpt-6.1-sol, the jq -e check prints true, and on your test machine /status shows the managed source and Claude Code refuses a session on the provider you left off.

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.

You are a staff engineer helping me make three changed defaults explicit for my team. I will paste the output of codex --version from our machines, any ~/.codex/config.toml model lines, whether our developers use Codex on Windows, our Copilot plan and model policy settings for GPT-6.1 Sol, the provider each developer signs in to Claude Code with, and the contents of any managed-settings.json we already deploy. Using only what I paste, tell me which machines would fall back to the GPT-6.1 Sol catalog default and whether to pin a model, which Windows users could try Codex v0.159.2, whether our Copilot policy leaves the new model on or off, and which allowedProviders list would match the providers we use without locking anyone out, including whether it must sit in the managed file instead of only the admin console. Remind me that allowedProviders is not in Anthropic’s docs yet and that I should confirm it against current documentation before rollout. Keep your first reply to that four-part checklist, do not invent settings, flags or versions I did not give you, and ask for anything missing before you expand.