OpenAI and Anthropic pushed out competing command-line coding agent updates within 48 hours of each other this week, and the timing was not a coincidence. On October 9, 2026, Anthropic shipped Claude Code 2.1.296, adding a workflow-level model control that lets teams pin every subagent in a pipeline to one chosen model. Days earlier, OpenAI’s Codex CLI crossed from version 0.161.0 into 0.162.0, adding managed Git worktree tools and task pinning inside its Command Center interface, after already switching its default model to the newly released GPT-6.1 Sol.
Neither release is flashy on its own. Taken together, they mark the point where terminal-based coding agents stopped being a novelty bolted onto a chat model and started shipping the kind of workflow plumbing that enterprise engineering teams actually ask for: isolated working directories, model-selection policy, and audit-friendly session controls. For developers deciding which Codex CLI or Claude Code build to standardize on, the differences in this release cycle say a lot about where both companies think the next year of agentic coding is headed.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Codex CLI 0.162.0: What OpenAI Actually Shipped
According to OpenAI’s own developer changelog, the Codex CLI update that shipped as version 0.161.0 made GPT-6.1 Sol the tool’s default model and added the ability to sign in to remote MCP servers directly from the terminal, extending how the CLI reaches external tools and context providers. The follow-up 0.162.0 build, tracked by the AI release-monitoring site AI/TLDR, went further: it introduced managed Git worktree tools for trusted local projects, added task pinning inside Codex’s Command Center panel, shipped a /copy command for grabbing transcript blocks, and fixed an underlying patch-application bug so CRLF line endings survive edits on Windows-based repositories.
The Git worktree addition is the one developers are watching most closely. Worktrees let a single repository check out multiple branches into separate directories at once, which is exactly the kind of isolation an autonomous coding agent needs when it is running several tasks against the same codebase in parallel. Managed inside Codex CLI, the feature effectively gives the agent its own sandboxed branch to experiment in without a developer needing to juggle git worktree add commands by hand.
Claude Code 2.1.296: Anthropic Answers With Workflow Controls
Anthropic’s release lands on a different problem. Claude Code 2.1.296, which shipped October 9 according to AI/TLDR’s release tracker, introduced an environment variable called CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL that forces every subagent inside a multi-step workflow to run on one selected model, rather than letting each subagent default independently. For teams running cost-sensitive pipelines, that single switch turns model selection from a per-agent guessing game into a workflow-level policy decision.
The same release added per-subagent autoCompactWindow settings, giving individual subagents their own context-compaction thresholds instead of inheriting one global value, and patched a managed-hook bug in which a hook that denied a tool call was not actually ending the agent’s turn. It is a narrower, more surgical update than OpenAI’s, focused on making large multi-agent workflows predictable rather than adding new surface area.
export CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL=claude-sonnet-5-5
claude-code run ./workflows/release-pipeline.yamlThat pattern, pinning every subagent in a defined workflow to a single named model via an environment variable, is the kind of control that shows up when a vendor is responding directly to enterprise feedback about unpredictable token spend, rather than chasing a new feature category.
GPT-6.1 Sol Becomes Codex CLI’s Default Engine
The model sitting underneath Codex CLI’s new default is itself only a few weeks old. OpenAI’s changelog lists GPT-6.1 Sol as releasing on September 29, 2026, positioned for complex coding and professional work at a lower cost than the earlier GPT-6 Astra model. A DevDay recap published by the developer outlet Developers Digest described GPT-6.1 Sol as reaching intelligence close to Astra on agentic coding tasks while costing roughly one-fifth as much, with cached input tokens discounted further still. The developer-focused newsletter ThursdAI reported the same cached-input discount figure at 95 percent off standard pricing.
OpenAI also added an Ultrafast mode for GPT-6.1 Sol inside its Responses API, according to the official changelog, though access to that speed tier was gated to higher-volume subscription plans at launch. The practical effect for anyone running Codex CLI day to day: the default model swap is also a default price cut, which matters once an agent is running dozens of tool calls per coding session rather than answering a single prompt.
Feature Snapshot: Three Coding Agents, One Week of Releases
The table below lines up the two headline releases against JetBrains’ own October entry into the space, the open-weight Mellum 2.1 model, to show how differently each vendor is approaching the same basic problem of running code-editing agents from outside a full IDE.
| Tool | Latest Version | Release Date | Core Model | Headline October Feature | Open Weight |
|---|---|---|---|---|---|
| OpenAI Codex CLI | 0.162.0 | Oct 9-11, 2026 | GPT-6.1 Sol | Managed Git worktree tools, Command Center task pinning | No |
| Anthropic Claude Code | 2.1.296 | Oct 9, 2026 | Claude Opus/Sonnet/Haiku 5.5 family | Workflow-level subagent model pinning | No |
| JetBrains Mellum | 2.1 | Oct 2026 | Mellum (12B, mixture-of-experts) | Open-weight release aimed at coding agents | Yes |
JetBrains took the opposite route from OpenAI and Anthropic: instead of expanding a closed terminal agent, it released Mellum 2.1 as an open 12-billion-parameter mixture-of-experts model specifically aimed at coding-agent workloads, according to the AI release tracker maintained by nocode.mba. That gives teams who want to self-host a coding model a credible option that did not exist a year ago, even as the two market leaders keep iterating on closed, subscription-priced CLIs.
Why the CLI Battlefield Matters Right Now
Command-line coding agents are no longer a side project for either company. Codex CLI and Claude Code both sit downstream of billions of dollars in model development, and both vendors are now shipping incremental CLI releases on a roughly weekly cadence, according to release logs tracked by AI/TLDR. That pace mirrors how browsers used to ship point releases during the early 2010s browser wars: fast, frequent, feature-by-feature, with each vendor watching the other’s changelog closely.
The stakes are also no longer theoretical. Coding agents have become the most visible front of the broader agentic-AI push, and adoption data backs that up. According to the 2026 IDC Future Enterprise Resiliency and Spending Survey, 74 percent of organizations with 500 or more employees had deployed at least one AI agent. A separate IDC analysis on agentic AI budgets, presented in September 2026, found that coding agents accounted for 53.6 percent of organizations’ agentic-AI budget allocations among the agent categories it measured, well ahead of other agent types like customer-service or data-analysis bots.
The Developer Survey Numbers Behind the Rush
Developer-side adoption figures tell a similar story. In the 2026 Stack Overflow Developer Survey, 65.9 percent of respondents reported using AI coding assistants or coding agents, up sharply from the prior year, when the 2025 Stack Overflow Developer Survey found 84 percent of respondents said they were using or planning to use AI tools in their development process. The gap between “using” and “planning to use” in that earlier survey has mostly closed over the past twelve months as CLI-based agents matured past their 2024-era autocomplete roots.
Within the coding-agent category specifically, the 2026 survey data shows a clear leader. Stack Overflow’s breakdown of AI code agent usage found Claude Code was used by 65.5 percent of respondents who reported using any of the listed AI coding agents, compared with 29.5 percent for OpenAI Codex. That gap is a big part of why OpenAI’s Codex CLI releases have picked up pace this quarter: every new feature is also an attempt to close a usage gap that Anthropic currently leads by more than two to one.
Pricing: What Each Model Actually Costs to Run
Model pricing underneath these CLIs has become its own competitive axis. Anthropic’s Claude Haiku 5.5, released October 7 and built for high-volume work like summarization, subagent orchestration, browser use, and lightweight coding tasks, lists at $0.10 per million input tokens and $0.50 per million output tokens, according to the same release-tracking coverage cited by letsdatascience.com’s LLM news desk, with support for prompts up to 100,000 tokens and configurable effort levels. That release followed Claude Opus 5.5 on September 22 and Claude Sonnet 5.5 on September 28, rounding out Anthropic’s full 5.5 model family in under three weeks.
| Model | Vendor | Release Date | Input Price (per 1M tokens) | Output Price (per 1M tokens) |
|---|---|---|---|---|
| Claude Haiku 5.5 | Anthropic | Oct 7, 2026 | $0.10 | $0.50 |
| Claude Sonnet 5.5 | Anthropic | Sep 28, 2026 | $2.00 | $10.00 |
| Claude Opus 5.5 | Anthropic | Sep 22, 2026 | $4.00 | $20.00 |
| GPT-6.1 Sol | OpenAI | Sep 29, 2026 | ~1/5 of GPT-6 Astra; cached input 95% off | ~1/5 of GPT-6 Astra |
OpenAI has not published an absolute per-token rate for GPT-6.1 Sol as clean as Anthropic’s published tiers; its DevDay messaging has stuck to relative framing against GPT-6 Astra. For engineering leads trying to budget a coding-agent rollout, that asymmetry alone is a reason teams keep both CLIs in rotation rather than standardizing on one: Anthropic’s pricing is easier to model into a monthly token budget, while OpenAI is betting that a 95-percent cached-input discount will win on real-world agent workloads where the same repository context gets re-read dozens of times per session.
From Autocomplete to Autonomous Agents: A Short History
It is easy to forget how recent all of this is. GitHub Copilot’s original autocomplete-style suggestions reached general availability in 2022, and for roughly two years the entire category meant “suggest the next few lines while I type.” Full command-line agents that could read a repository, plan multi-step changes, run tests, and open pull requests on their own did not become mainstream developer tools until 2025. Codex CLI and Claude Code both trace their lineage to that shift: terminal-native tools built to operate with minimal supervision across an entire coding session rather than one keystroke at a time.
What is different about this week’s releases is the subject matter. Early coding-agent updates were almost entirely about model quality: smarter completions, better bug fixes, more reliable test generation. The October 2026 releases are about operations, Git worktree isolation, workflow-level model pinning, context-window budgeting per subagent. That is the signature of a category moving from “can it code” to “can it run inside a real engineering organization’s existing processes.”
The Git Worktree Question: Convenience Versus Blast Radius
Managed Git worktree tools inside Codex CLI solve a real annoyance, letting an agent check out an isolated branch without a developer manually running worktree commands, but they also widen what an unsupervised agent can touch inside a local filesystem. OpenAI has scoped the feature to “trusted local projects,” according to the AI/TLDR release notes, a phrasing that puts the responsibility for defining trust back on the team running the CLI rather than on the tool itself.
That tradeoff is not unique to Codex CLI. Any agent given direct write access to a repository, whether through managed worktrees, direct branch commits, or automated pull requests, inherits the same basic risk: a bad instruction or a misread prompt can produce changes across more of a codebase than a human reviewer catches before merge. Engineering teams adopting either tool are increasingly treating agent-initiated branches the same way they would treat a junior contributor’s first pull request, with mandatory review gates rather than auto-merge.
Where JetBrains, Microsoft, and the Open-Source Side Fit In
OpenAI and Anthropic are not the only ones moving. Microsoft Research Asia released Agent Lightning version 1.0 on October 7, 2026, an open-source reinforcement learning framework for training AI agents using existing deployment harnesses, according to coverage from Agentry Newsroom. That release targets a different layer of the stack entirely, letting teams fine-tune agent behavior against their own infrastructure rather than relying solely on whatever a vendor’s latest model happens to do out of the box.
JetBrains’ Mellum 2.1, mentioned above as an open 12-billion-parameter mixture-of-experts model, points the same direction: a credible, self-hostable alternative to the Codex CLI and Claude Code duopoly for teams unwilling to send proprietary source code through a closed vendor API. Neither Agent Lightning nor Mellum has anywhere near the usage share Stack Overflow’s survey recorded for Claude Code or Codex, but both signal that the open side of the coding-agent market is no longer purely theoretical.
Market Impact: What This Release Cycle Means for OpenAI and Anthropic
For OpenAI, the push to make GPT-6.1 Sol the Codex CLI default while discounting cached input tokens looks like a direct attempt to close the usage gap Stack Overflow measured against Claude Code. Price is one of the few levers OpenAI can pull quickly against an incumbent that currently holds more than double its share among surveyed coding-agent users. For Anthropic, shipping workflow-level subagent model controls just two days after its own Claude Haiku 5.5 launch suggests a strategy built around giving enterprise customers more granular cost control rather than competing purely on raw model capability, where the two companies are reportedly close on agentic coding benchmarks according to the DevDay recap coverage.
The 53.6 percent of agentic-AI budgets IDC found going specifically to coding agents means neither company can afford to let its CLI fall behind on either price or workflow ergonomics. That budget concentration is also why rivals like JetBrains and Microsoft Research are investing in the space even without a dominant share: coding agents are where the agentic-AI money is actually being spent today, not a side bet on a category that might matter later.
Five Predictions for the Next Release Cycle
- Expect OpenAI to publish an absolute per-token price for GPT-6.1 Sol within the next quarter rather than relying solely on relative comparisons to GPT-6 Astra, as enterprise buyers push for budget-line clarity.
- Claude Code’s workflow-level model pinning will likely be followed by similar policy controls in Codex CLI, since both vendors have a pattern of matching each other’s enterprise-facing features within one or two release cycles.
- Git worktree-style isolation features will spread to other agentic coding tools as the default way to sandbox autonomous changes, rather than remaining a Codex CLI-specific feature.
- Open-weight entrants like JetBrains’ Mellum family will keep gaining ground in regulated industries where sending source code to a closed API is a compliance blocker, even if their overall usage share stays well below Claude Code and Codex.
- The usage gap Stack Overflow measured between Claude Code and OpenAI Codex is likely to narrow somewhat in the 2027 survey cycle as GPT-6.1 Sol’s pricing advantage pulls cost-sensitive teams toward Codex CLI, though a full reversal of the current 65.5-to-29.5 split looks unlikely within a single year.
What Teams Evaluating a Coding Agent Should Watch
For an engineering team deciding where to standardize, the practical questions this release cycle raises are less about raw capability and more about operations: does the workflow need per-agent model pinning for cost control, does the team’s branch-protection policy already account for an agent-managed worktree, and does procurement need an absolute per-token rate card or can it live with relative pricing claims. Those are the same questions that determine whether a tool survives a pilot and gets rolled out company-wide, regardless of which vendor’s benchmark numbers look better on a given week.
Frequently Asked Questions
What is new in Codex CLI 0.162.0?
Codex CLI 0.162.0 adds managed Git worktree tools for trusted local projects, task pinning inside the Command Center interface, a /copy command for transcript blocks, and a fix that preserves CRLF line endings when applying patches on Windows-based projects, according to release notes tracked by AI/TLDR.
What changed in Claude Code 2.1.296?
The release added the CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL environment variable so every subagent in a workflow can be pinned to one model, introduced per-subagent autoCompactWindow settings, and fixed a managed-hook bug where a denied tool call did not end the agent’s turn.
Is GPT-6.1 Sol now the default model in Codex CLI?
Yes. OpenAI’s changelog confirms Codex CLI switched its default model to GPT-6.1 Sol, which the company says delivers intelligence close to GPT-6 Astra on agentic coding tasks at roughly one-fifth the price, with a 95 percent discount on cached input tokens.
How much does Claude Haiku 5.5 cost to run?
Claude Haiku 5.5 lists at $0.10 per million input tokens and $0.50 per million output tokens, supports prompts up to 100,000 tokens, and offers configurable effort levels for tuning cost against response depth.
Which coding agent do more developers actually use, Claude Code or Codex?
Stack Overflow’s 2026 Developer Survey found Claude Code was used by 65.5 percent of respondents who reported using a listed AI coding agent, versus 29.5 percent for OpenAI Codex.
What is JetBrains Mellum 2.1?
Mellum 2.1 is an open-weight, 12-billion-parameter mixture-of-experts model JetBrains released in October 2026, aimed specifically at coding-agent workloads and offered as a self-hostable alternative to closed CLI tools like Codex CLI and Claude Code.
Are managed Git worktrees in Codex CLI safe to use on any repository?
OpenAI has scoped the feature to “trusted local projects,” meaning teams are expected to apply their own judgment about which repositories an autonomous agent should be allowed to branch and modify without added review gates.
How much of enterprise agentic-AI spending goes toward coding agents?
IDC’s 2026 Agentic AI at Scale research found coding agents accounted for 53.6 percent of organizations’ agentic-AI budget allocations among the agent categories it measured, ahead of other categories such as customer-service or data-analysis agents.
Related Coverage
- Claude Code vs Codex vs Copilot: 65.5% Dev Share [2026]
- Set Up Cursor With Claude Code: 13 Steps, 1M Tokens [2026]
- OpenCode vs Codex CLI vs Gemini CLI: 103K Star Gap [2026]
- GitHub Copilot CLI Adds Computer Use, 3 Rivals Chase [2026]
- Zed Adds BYOK for Claude Opus 5.5, GPT-6.1 Sol [2026]
