OpenAI shipped Codex CLI 0.137.0 on June 4, 2026, adding multi-agent v2 and enterprise credit controls. With Anthropic’s Agent SDK billing change scheduled for June 15 and OpenAI offering two free months to enterprises switching off Claude Code, the June release cadence was a switching-cost maneuver, not a product roadmap. [Updated June 2026] The cadence held; the deadline did not. Codex reached 0.142.2 by June 25, including an /import command that pulls Claude Code config and chat history straight into Codex. Anthropic, meanwhile, paused the June 15 billing change before it took effect, which removes the forcing function the whole maneuver was timed against.
The 45-minute counter and what triggered it
According to Ramp’s April 2026 AI spending index, Anthropic reached 34.4% of enterprise AI spend versus OpenAI’s 32.3%, the first time Anthropic held the lead. The original Ramp report was not independently fetched for this analysis; the figures come from a secondary Chinese-language summary and should be read as directional rather than precise.
OpenAI launched a two-free-months enterprise promotion on May 13, 2026, offering net-new Codex seats to teams switching from Claude Code within a 30-day window. OpenAI had already dropped the seat fee for Codex’s business tier the prior month, stacking the promo on a lower baseline price.
Anthropic’s counter came within 45 minutes: a 50% increase in Claude Code weekly usage limits, running through July 13, 2026. The two moves create a narrow window where both vendors are subsidizing usage to prevent defection, and that window is bounded by Anthropic’s Agent SDK billing change on June 15.
What Codex 0.137.0 ships
The June 4, 2026 release of Codex CLI 0.137.0 adds multi-agent v2 with per-thread runtime persistence, parallel standalone web search, enterprise monthly credit limits displayed in the status bar, cloud-managed configuration bundles for EDU workspaces, and a skills extension scaffold for custom tool integrations.
Per-thread runtime persistence means multi-agent sessions survive process restarts. That is the specific capability gap that made agent-based workflows on Codex fragile compared to Claude Code’s session continuity. Cloud-managed config bundles address a different switching cost: EDU administrators previously had to configure Codex environments per-seat, per-machine. Centralized config removes that operational barrier.
The prior release, Codex CLI 0.136.0, added session archiving (codex archive/unarchive), CODEX_API_KEY authentication for remote exec-server registration, a feature-gated standalone image generation pipeline, a Python SDK beta (pip install openai-codex), and command-safety hardening that blocks /diff from running repo-provided Git hooks. The /diff hook block closes a targeted attack surface: a malicious pre-commit hook in a cloned repository could execute arbitrary code during a Codex-assisted review. Blocking repo-provided hooks during /diff prevents that.
Two other June releases round out the surface. Codex Sites, announced in June 2026, is a plugin for creating, deploying, and inspecting websites hosted by OpenAI, with ChatGPT Business including Sites by default and Enterprise admins controlling access via RBAC. Amazon Bedrock support, announced June 1, 2026, allows local Codex execution with AWS-managed authentication and billing, falling back to AWS_REGION/AWS_DEFAULT_REGION environment variables.
What shipped after 0.137.0 [Updated June 2026]
The release cadence the original analysis flagged as unsustainable kept going anyway. By June 25 the CLI had moved through seven point releases to 0.142.2, and the pattern is consistent: most of what shipped lowers the cost of leaving Claude Code rather than adding capability Claude Code lacks.
The clearest example is the /import command in 0.140.0, which landed June 15 and selectively imports setup, project configuration, and recent chats from Claude Code. A team’s CLAUDE.md context, slash-command setup, and in-progress conversations no longer have to be rebuilt by hand. That is the most direct switching-cost cut OpenAI has made, because it targets the one barrier the promo and the free seats do not: the accumulated local configuration that makes an incumbent tool sticky. The date is not a coincidence. 0.140.0 shipped the same day Anthropic’s billing change was scheduled to bite.
Three other releases move the comparison:
- 0.139.0 (June 12) put web search directly into Code mode and added an indexed web-search mode that allows live queries while restricting page access to server-approved URLs. That is a data-governance control for enterprises, not a developer convenience.
- 0.141.0 (June 19) rebuilt remote execution on authenticated, end-to-end encrypted Noise relay channels, making the shared-team-server model (the
CODEX_API_KEYexec-server registration from 0.136.0) safer to run across a team. - 0.142.0 (June 23) added usage-credit redemption inside
/usageand configurable multi-agent delegation an admin can set to disabled, explicit-request-only, or proactive at the thread and turn level. Delegation policy is the admin-facing half of the multi-agent v2 work 0.137.0 started.
None of this is differentiation. It is the methodical removal of every operational reason a team gives for staying: config migration, governance controls, team-server topology, admin policy. The skills extension scaffold and Codex Sites, flagged earlier as the actual differentiation bets, did not see comparable follow-through over the same three weeks.
| Lever | Status as of June 26, 2026 |
|---|---|
| Codex CLI version | 0.142.2 (was 0.137.0) |
| Claude Code config migration | /import shipped in 0.140.0 |
| OpenAI two-free-months promo | 30-day enrollment from May 13 closed mid-June |
| Anthropic Agent SDK credit pool | Paused before June 15; subscription billing unchanged |
| Claude Code 50% weekly-limit bump | Active through July 13, 2026 |
JetBrains integration and the IDE surface grab
In January 2026, JetBrains shipped native Codex integration into its AI chat, starting with IDE version 2025.3. The integration supports three authentication paths: a JetBrains AI subscription, a ChatGPT account, or a bring-your-own-key API key. A limited-time free promotion via JetBrains AI ran from January 22, 2026.
As of June 2026, the Codex IDE extension covers VS Code, Cursor, Windsurf, Visual Studio Code Insiders, and JetBrains IDEs on macOS, Windows, and Linux. It includes multiple interaction modes with varying autonomy levels, cloud delegation, and image generation inside the editor.
This is where the switching-cost calculation gets concrete. A team on Claude Code that also uses JetBrains has a zero-configuration path to Codex inside the same IDE as of January 2026. The BYOK option means the switching cost for an individual developer to test Codex is one API key rotation, not a new subscription. OpenAI’s two-free-months promo reduces it further to zero for the first 60 days. The question is whether the evaluation happens before Anthropic’s June 15 billing change makes the comparison more expensive to run.
Anthropic’s Agent SDK credit pool: why June 15 matters
Anthropic’s Agent SDK credit pool was scheduled for June 15, 2026, moving programmatic Claude usage out of subscription quotas into a separate monthly credit allocation: $20 for Pro subscribers, $100 for Max 5x, and $200 for Max 20x subscribers, billed at standard API rates. Credits would not roll over. This was the same metered-billing direction Anthropic took for large enterprise seats when it ended flat-fee Claude above 150 seats.
[Updated June 2026] It did not arrive. On June 15 Anthropic paused the change before it took effect, telling subscribers there is no credit to claim and that Agent SDK, claude -p, GitHub Actions, and third-party app usage keep drawing from existing Pro, Max, Team, and Enterprise limits exactly as before. The company said it is reworking the plan and will give advance notice before any revised version ships. The deadline this section was built around evaporated. That removes the forcing function, not the underlying question: a metered Agent SDK pool is still Anthropic’s stated direction, now with no date attached.
The original thesis was that this deadline shaped the switching-cost window. Teams that use Claude Code for agent-based workflows (multi-step refactors, test generation, codebase-wide analysis) would see those workloads move from their existing subscription pool to a metered credit system. If agent usage exceeded the credit cap, the non-rollover policy created a direct incentive to evaluate whether Codex’s pricing was cheaper for that workload class, especially while the free-months promo was active. The pause neutralizes that incentive for now.
[Updated June 2026] With the change shelved, the only signal that ever materialized was developer reaction to the announcement; the credit pool never went live. Near-term pressure on Claude Code agent workloads now comes from the weekly usage limits, not a separate metered pool.
A switching-cost audit for teams
The practical decision for teams on Claude Code is not “which tool is better” but “what does it cost to find out.” Both vendors are subsidizing the comparison. The parity window is time-limited.
A decision framework:
Map current Claude Code usage patterns. Identify which workloads use agentic features (multi-file edits, multi-step reasoning, autonomous refactors). These were the workloads the paused Agent SDK credit pool would have metered, and they remain the ones most exposed if Anthropic revives the plan. [Updated June 2026]
Test Codex on the same workloads while the entry path is cheap. OpenAI’s two-free-months offer was open to enterprises switching from Claude Code within 30 days of May 13, 2026. [Updated June 2026] That enrollment window closed in mid-June, so teams that did not apply lost the free-months entry point; the lower business-tier baseline, the JetBrains integration, the Bedrock auth, and the new
/importmigration command still cut setup cost for teams already in those ecosystems.Compare real cost, not list cost. Factor in Anthropic’s credit cap and non-rollover policy against Codex’s seat pricing (with the business-tier fee reduction already applied as of May 2026). Model worst-case months, not average months.
Evaluate lock-in vectors on both sides. Codex’s cloud-managed config bundles, Bedrock billing integration, and Sites hosting each add dependencies that increase the cost of switching away later. Claude Code’s lock-in vectors include the Agent SDK credit system and Anthropic-specific prompt patterns baked into existing workflows.
Set a decision date before July 13. Anthropic’s 50% usage-limit bump expires then. After that, the subsidized comparison window closes and both vendors revert to baseline pricing.
What to watch
Multiple June 2026 Codex releases (Bedrock support on June 1, CLI 0.137.0 on June 4, and Codex Sites) compressed switching-cost reduction into a single week. The original read was that the pace would not sustain. It did: seven more point releases landed through 0.142.2 on June 25, and the Claude Code /import command arrived in the middle of them. [Updated June 2026] The parity features are shipping because competitive pressure from Anthropic’s spend lead, as reported by Ramp’s April 2026 index, is real enough to justify the engineering sprint.
For practitioners, the actionable signal is the gap between parity features and differentiated features. Multi-agent v2, Bedrock auth, and IDE coverage are parity plays that close existing gaps. The skills extension scaffold in 0.137.0 and the Codex Sites hosting model are differentiation attempts. Whether either becomes a genuine lock-in vector depends on adoption rates that are not yet measurable as of June 4, 2026.
The June 15 comparison point did not happen the way it was supposed to. [Updated June 2026] Anthropic shelved the Agent SDK credit pool rather than letting teams observe live metering, which buys Claude Code time but leaves the metered direction on the table without a date. The remaining hard date is July 13, when the 50% usage-limit bump expires and Claude Code reverts to baseline weekly limits. The next signal worth watching is whether Anthropic reissues the credit plan with advance notice, as it promised, or whether OpenAI’s steady stream of migration-cost cuts (the /import path most of all) pulls enough teams across that the question answers itself. The vendor that has established more friction in its exit path wins the renewal.
Frequently Asked Questions
Does Codex have a programmatic SDK comparable to Anthropic’s Agent SDK?
Yes. Codex CLI 0.136.0 shipped a Python SDK beta (pip install openai-codex) for scripting Codex from CI/CD pipelines or batch workflows, so teams evaluating both sides have a working programmatic surface on the Codex side. [Updated June 2026] Anthropic’s Agent SDK is also live; the June 15 event was a billing change, not a launch, and it moved Agent SDK usage to a separate metered credit pool. Anthropic paused that change before it took effect, so Agent SDK usage still draws from subscription limits as before.
Can Codex-on-Bedrock spend count toward existing AWS commit discounts?
Routing Codex through Amazon Bedrock puts the compute charges on the regular AWS bill, where they count toward Enterprise Discount Program or Private Pricing Agreement thresholds. Teams already above their committed spend floor pay a marginal cost below Codex list price. This is the same procurement advantage Anthropic uses with its cloud partnerships, but applied to the OpenAI side.
Is the skills extension scaffold in 0.137.0 ready to influence a switching decision?
Not as of June 4. The scaffold shipped as an extension point for custom tool integrations, but no published documentation, example skills, or interface contracts exist yet. It signals OpenAI’s direction for extensibility but should be treated as forward-looking, not a factor in this month’s vendor comparison.
Which current Codex capabilities have no Claude Code equivalent?
Two that stand out: the feature-gated standalone image generation pipeline (from 0.136.0, also available inside the IDE extension) and the Codex Sites hosting plugin for deploying web apps from the CLI. Claude Code has neither image output nor a hosted deployment path. The CODEX_API_KEY remote exec-server registration also enables a shared-team-server model where one Codex instance serves multiple developers, a deployment pattern Claude Code’s local-first architecture does not support.
