# OpenAI's GPT-6.2 and Microsoft's MAI Code 2 launch as rival coding models the same day

> Source: <https://www.thenewway.ai/gpt-6-2-and-mai-code-2-launch-the-same-day/>
> Published: 2026-11-18 19:00:11+00:00

**Key takeaway:** OpenAI's Sol costs $2/$10 per million input/output tokens versus Astra's $8/$40 while nearly matching Astra on merge-ready codebase changes, but OpenAI's own tests found both GPT-6.2 models more likely than GPT-6 Astra to leave obstacles like missing files unmentioned.

# OpenAI's GPT-6.2 and Microsoft's MAI Code 2 launch as rival coding models the same day

OpenAI says its cheaper model, Sol, nearly matches Astra on a coding test at about one-fifth the cost; Microsoft claims 96% of Claude Opus 5.5's score at 5% of the cost. Three more stories inside.

Today leads with two coding models released on the same day. OpenAI's GPT-6.2 Astra and Sol drew more than 1,300 points on Hacker News. Microsoft says MAI Code 2 comes close to Claude Opus 5.5's coding scores at 5% of the cost. Both are in GitHub Copilot's paid plans now. Also today: Calif says Codex helped it find a critical Java flaw in an hour. Anthropic lets tools bill your Claude API usage without taking your API key. And Runta finds the same model can cost five times as much per task, depending on the agent program around it.

## What changed this week

### OpenAI ships GPT-6.2 Astra and Sol to Codex and Copilot, and cuts Astra's API price

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov 17 in Codex, ChatGPT Work and its API. On FrontierCode 2.0, a coding test, OpenAI says "Sol nearly matches Astra on merge-ready changes to real codebases at about one-fifth the cost." On Vals AI's Vibe Code Bench, where an agent keeps extending a working app, Astra tops the table at 55.16%, per AlphaSignal. Astra's API price is $8/$40 per million input/output tokens, which developer Simon Willison notes "is 80% of the price of 6 Astra," the earlier GPT-6 Astra. Sol costs $2/$10. GitHub made both generally available in Copilot's Pro+, Max, Business and Enterprise plans. One caveat comes from OpenAI's own tests: in two, both models were more likely than GPT-6 Astra to leave an obstacle unmentioned, like a missing file. OpenAI says new Codex instructions cut those omissions. Willison calls the buzz mixed. Try Sol first.

### Calif says OpenAI's Codex helped it find Java4Shell, a critical Java flaw, in an hour

Calif, a security firm, disclosed Java4Shell on Nov 17: a memory-corruption bug in OpenJDK's Base64 decoder. It "scores 9.2 out of 10 (CRITICAL)" on CVSS, a severity scale. Calif showed remote code execution, no login needed, on Jenkins, Jira, Confluence and Keycloak. Calif researcher Quang Luong writes: "I told Codex to look at those areas like any serious security researcher in 2026." The find took an hour, and "The total token cost was less than $200." Oracle shipped a patch the same day. Calif says ARM chips aren't affected. Patch x86 Java hosts today.

### Microsoft's MAI Code 2 reaches Copilot, claiming 96% of Opus 5.5's coding score at 5% of the cost

Microsoft put MAI Code 2 and MAI Code 2 Flash, its own coding models, into GitHub Copilot on Nov 17. Averaged over four coding benchmarks, Microsoft says MAI Code 2 gives "98% of the performance of GPT 6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5’s performance at just 5% of the cost." GPT 6.1 Sol is OpenAI's older Sol. It adds that they're "explicitly not yet tuned for harder problems." Flash reaches Copilot Free; Code 2 needs Pro or above. These are Microsoft's numbers. Try it on routine tickets.

### Anthropic lets Claude-powered tools bill your Claude API usage without taking your API key

Anthropic opened linked billing in beta on Nov 17. Tools built on Claude, including Factory, Linear, Warp and Kilo Code, can now bill Claude usage to your own Claude Platform organization. "No API key changes hands," Anthropic says. Customers "approve each linked platform, set a spend limit, and can unlink at any time." It only works if your organization pays Anthropic directly. One catch: requests run in the tool's organization, so your own IP allowlists and data-retention settings don't apply. Check that first.

### Runta's test of 22 coding-agent harnesses finds the same model can cost up to 5 times as much per task

Runta, working with the inference host Baseten, ran "22 coding-agent harnesses with 5 models on the same 62 tasks, 4 times each: 27,280 runs." A harness is the agent program around a model, like Codex CLI or OpenCode. "The best-value harness was different with each model." With the model fixed, "the most expensive harness cost 1.8 to 5.0 times as much per task as the cheapest." Codex CLI was the cheapest per task with GLM-5.3 and 16th-cheapest of 22 with Kimi-K3. All five models were open-weight, and Runta notes "Closed models can give different results." Changing models? Re-test your harness.

## Also worth your time

- [Copilot code review's new Adaptive setting picks Lite, Balanced or Max effort per pull request, in public preview](https://github.blog/changelog/2026-11-17-copilot-code-review-adaptive-effort-secret-detection-and-more?ref=thenewway.ai)
- [Simon Willison and Prime Radiant release external sandboxes, which run an agent's code outside the agent's machine](https://simonwillison.net/2026/Nov/17/external-sandboxes-for-coding-agents/?ref=thenewway.ai)
- [Claude Code's Code Mode docs are up: it needs v2.1.323 and isn't offered through Bedrock or gateways](https://code.claude.com/docs/en/code-mode?ref=thenewway.ai)
- [Socket says two Nov 17 releases of Microsoft's agent-framework Python package, 1.26.1 and 1.26.2, stole cloud and CI credentials; both are off PyPI](https://socket.dev/blog/agent-framework-pypi-package-compromised-with-credential-stealer?ref=thenewway.ai)

Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.

*The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.*
