Disclosure: TechSifted has no affiliate relationship with OpenAI, Google, or Microsoft. This is editorial news coverage.

Three things caught my attention in the past 48 hours. OpenAI’s latest model update quietly breaks a pattern. Google filled the gap in its own lineup with a model that was arguably missing. And Microsoft showed up at Ignite with coding models that are, at minimum, a credible challenge to OpenAI and Anthropic inside GitHub Copilot.

1. OpenAI GPT-6.2 Astra and Sol: Better Models, Lower Price Tags

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17. Two models at once, both positioned as upgrades over their direct predecessors.

GPT-6.2 Astra is the new frontier model. Its API pricing comes in at $8 per million input tokens, $0.40 per million cached, and $40 per million output. That is 20% cheaper on input and output than GPT-6 Astra. On Agents’ Last Exam, a benchmark that tests multistep professional work, GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra. It also scores 5.5 points ahead of Claude Fable 5.2 on the same benchmark, at 5% of the cost.

GPT-6.2 Sol is the everyday model. It keeps GPT-6.1 Sol’s API pricing. On Cognition’s FrontierCode 2.0, which tests whether changes to real codebases meet maintainers’ standards for merging, Sol comes within 0.5 points of GPT-6.2 Astra at roughly one-fifth the cost. OpenAI says both models deliver better results with fewer output tokens, so you can get more done within your usage limits.

OpenAI is also calling Astra its best model for scientists.

Both models are available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, and through the OpenAI API.

Why it matters. The consistent story here is efficiency: more intelligence, lower cost, fewer tokens per task. That is a real improvement, not a rebranding. The 5% cost claim relative to Fable 5.2 on that one benchmark is worth verifying on your own workloads before you read too much into it. OpenAI’s own footnote says the Fable 5.2 leaderboard results exclude nine tasks that used a fallback model. But if the direction holds across categories, the gap between frontier models and what businesses can afford to run at scale is narrowing faster than I expected.

Find out more: Introducing GPT-6.2 Astra and Sol (OpenAI)


2. Google Gemini 4 Neon: The Middle Model Gets Filled In

Google released Gemini 4 Neon today, November 18. It sits between Hydrogen (fast everyday execution) and Carbon (the hardest, long-horizon projects) in the Gemini 4 family.

Gemini 4 Neon sign with the Google Gemini star logo in orange neon lighting

Image: Google

Neon is priced at $3 per million input tokens, $0.15 per million cached (a 95% discount on cached tokens), and $15 per million output tokens. Carbon and Argon are on a promotional rate of $2/$10 per million through November 30, so Neon is more expensive right now. Starting December 1, Carbon’s rate goes to $4/$20, at which point Neon would be 25% cheaper than Carbon per token.

The benchmarks Google is pushing: Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level (Source). It ranks second on steerability on the Agent Arena Index, meaning it handles mid-task user corrections better than most. It hits first on Legal Research Bench, Tax Agent Bench, and Zapier’s AutomationBench (56.3%). On FrontierCode 2.0 at its default high thinking setting, 46.2%.

The model has a 1-million-token context window and four thinking levels: Low, Medium, High (default), and xhigh (Source). It is available in the Gemini API via Google AI Studio and in Gemini for business, and for paid Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app.

Google says Neon costs 32% less per task than Carbon at the default high thinking level.

Why it matters. The Gemini 4 family was hard to explain to anyone not deep in AI tooling. Carbon for hard things, Hydrogen for fast things, Argon for coding and analysis. Neon simplifies the lineup by giving teams something Google pitches as faster and cheaper per task than Carbon. Whether that holds in practice depends on the workload. The Legal Research and Tax Agent results are interesting if you work in those fields. The generic benchmark numbers, less so.

If you have been using Carbon for routine coding or analytical tasks, the December pricing shift is worth watching. At the new rates, Neon would be a meaningfully cheaper option for a lot of that work.

Find out more: Introducing Gemini 4 Neon (Google)


3. Microsoft Launches Its Own Coding Models at Ignite 2026

At Ignite 2026 on November 17, Microsoft launched MAI Code 2 and MAI Code 2 Flash, two in-house coding models now available in GitHub Copilot.

MAI Code 2 is designed for the kind of coding work enterprise teams actually do every day: deep debugging, changes across multiple files, repository-wide reasoning, and UI development. MAI Code 2 Flash handles high-volume tasks: code explanations, targeted edits, routine fixes, and rapid iterations. Microsoft is explicit that neither model is yet tuned for harder problems like very long-horizon agentic work or fundamental science research (Source).

The models are close in performance. The gap Microsoft is emphasizing is cost.

Averaged over four benchmarks aimed at the everyday enterprise coding tasks Microsoft sees in GitHub Copilot, MAI Code 2 delivers 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of the cost (Source). Flash follows the same pattern. Microsoft says it offers comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost, respectively (Source).

At promotional pricing, good until May 17, 2027, MAI Code 2 runs $0.50 per million input tokens and $2.50 per million output (Source). Post-promotion it goes to $1.00 and $5.00. Flash is $0.05 input, $0.25 output now, and $0.10 and $0.50 after the promotion ends.

Worth knowing: the cost comparisons Microsoft published use the 50% promotional pricing for the MAI models. That gap narrows when the promotion ends.

For more context on GitHub Copilot pricing and how Microsoft positions itself in the coding tools market versus competitors like Cursor, those earlier pieces are still relevant background.

Why it matters. Microsoft is working to reduce its reliance on outside model providers, OpenAI and Anthropic included. Capable in-house coding models are a direct way to do that. The benchmarks suggest the trade-off is roughly equivalent performance for a lot less money, at least at current promotional rates. If you run GitHub Copilot at scale, this matters for your bill. If you do not, it is a signal that the tier of models OpenAI and Anthropic charge enterprise prices for is becoming easier to replicate in-house.

For more context on where these models fit against the GPT-6 Astra era launch, that piece from September covers what GPT-6 Astra changed about the competitive picture.

Find out more: Microsoft launches MAI Code 2 models (InfoWorld)