GPT-6.2’s biggest leap is in work per dollar

Welcome back.  Microsoft used its Ignite event in San Francisco on Tuesday to make a blunt pitch to enterprises tired of rising token bills: do more for less, and prove the ROI. OpenAI echoed similar themes with GPT-6.2 Astra and Sol, which top several benchmarks but whose bigger leap is how much work they deliver per dollar. Meanwhile, Mistral is steering away from that price war, building sovereign models for industrial engineering and other regulated sectors that want to own their intelligence outright. —Jason Hiner

IN TODAY’S NEWSLETTER

1. GPT-6.2 shifts the AI race toward cost per task

2. Microsoft challenges the price of intelligence

3. Mistral sees a market the AI giants may miss

PRODUCTS

GPT-6.2 shifts the AI race toward cost per task

New AI models have been leapfrogging each other nearly every week during 2026, and a new week is upon us.

On Tuesday, OpenAI announced GPT-6.2 Astra and Sol, its two most powerful models that are now racing past recently unveiled models from Anthropic and Google in several benchmarks. But the bigger story is around the GPT-6.2 advances in cost and efficiency.

OpenAI characterizes GPT-6.2 Astra as the model for the most challenging and complex professional tasks and workflows. That includes creating slides, spreadsheets, and docs, which are the vessels containing the work of so many professionals day-to-day. With Astra, OpenAI says the model will "more closely follow the templates you provide and the conventions of your field." The company also flagged writing cleaner code and building front-end web sites that work better and are more aesthetically pleasing as highlights of Astra's growing capabilities. 

Astra costs the same as Anthropic's Fable 5.2: $8 per million input tokens and $40 per million output tokens. However, Astra costs twice as much as Fable per million cached input tokens: $0.40 compared to $0.20. 

As for Sol, OpenAI emphasizes that it's the smart model focused on efficiency to let you complete more tasks with fewer tokens. It's especially aimed at agentic multistep workflows, writing that is cleaner and more natural, and collaborating with humans in an intuitive way. The AI lab revealed, "Many employees at OpenAI use it as their default model for both technical and non-technical tasks."

While Sol, which is OpenAI's second-tier model, offers roughly the equivalent intelligence as Anthropic's second-tier Opus model, Sol costs the same as Anthropic's third-tier Sonnet model: $2 per million input tokens, $10 per million output tokens, and $0.10 per million cached input tokens. 

Here's where both new models deliver in terms of efficiency and affordability:

  • Using the Agent's Last Exam benchmark that tests multistep professional workflows, GPT-6.2 Astra set a new high score (78.3%) while scoring 5.2 points higher than GPT-6 Astra at less than a third of the cost and 5.5 points higher than Fable 5.2 at 5% of the cost for the tokens to complete the work

  • Using Cognition’s FrontierCode 2.0 benchmark, GPT‑6.2 Sol scores only 0.5 points lower than GPT‑6.2 Astra at about one-fifth of the cost (both using the max setting); and both come in slightly lower than Claude Fable 5.2 and Claude Opus 5.5, which both cost considerable more

  • In OpenAI's GDPval v2 benchmark, which measures real-world, complex, long-horizon knowledge work from sectors such as marketing, finance, data science, and biotech, GPT-6.2 Astra leads all models tested, improves 6.2% from GPT-6, and uses 22% fewer output tokens

  • GPT-6.2 Astra also set new highs on Terminal‑Bench Science 0.1 and the Vals Science Index benchmarks, making it one of the top models in the world for use in science labs

I learned about the release of GPT-6.2 Astra and Sol in the Microsoft Ignite keynote this morning, shortly before the news became public. When Microsoft CEO Satya Nadella matter-of-factly named the models as part of the cutting edge AI available on Microsoft's platform, I immediately thought, "Wait, did those just get announced?" It's a sign of the times that the announcement of new models with better capabilities and the ability to complete more and better tasks with fewer tokens is just part of our everyday reality now. Of course, it's great to get models that can do more for less. But it also reminds me that the models themselves are getting more and more commoditized. Routing tasks and workloads to the right models to increase speed and lower costs is where a lot of the value is accruing right now. And while Anthropic played that up as one of the built-in capabilities of Fable 5.2, there's no sign of routing or orchestrating in GPT-6.2 Astra. That's where OpenAI will need to go next to better serve its users and avoid having to compete almost solely on price. 

Jason Hiner, Editor-in-Chief

SPONSORED

TOGETHER WITH ZENDESK

Outcomes are all people care about. And that’s the truth.

Service has spent decades obsessing over traditional metrics like response times and efficiency.

People today don’t just want any old response. They’re here for the destination (service outcome) not the journey (service interaction).

Zendesk's Outcome AI Platform connects intent to action, helping businesses move beyond helpful responses and efficiency to deliver meaningful outcomes.

The real measure of a great experience is whether you get the outcome you need.

BIG TECH

Microsoft challenges the price of intelligence

Microsoft Ignite included a waterfall of updates across the tech giant's portfolio, but many of them followed one central thread: Efficiency. 

Across a number of the model releases and new product offerings Microsoft CEO Satya Nadella announced at the Chase Center in San Francisco on Tuesday, the company highlighted cost efficiency as one of the biggest advantages of its suite, as enterprises largely grow fatigued of rising token bills and continue to question when they will see returns on their investments.

Based on a recent report from GigaOm, Microsoft claims that, for comparable quality, Copilot is 84% cheaper than Claude and 63% cheaper than ChatGPT for the average user. Power users, meanwhile, find Copilot to be 68% cheaper than Claude and 35% cheaper than ChatGPT. 

"Anything you do in Claude or ChatGPT, you can do in Copilot for less," Judson Althoff, CEO of Microsoft's commercial business, said in the keynote. 

This shift is particularly evident in its model announcements, debuting: 

  • Microsoft-Decision-1.1, a decision model similar to Jev and following in the footsteps of OpenAI and Google, which both recently released their own competing decision models. The company said that Microsoft-Decision-1.1, now generally available, is ten times faster and 200 times less expensive than traditional LLMs.

  • Additionally, the company said that Auto, its model routing service, will run on Microsoft-Decision-1.1. 

  • The company's other new models, including MAI-Cyber-2-Flash and MAI-Code-2, meanwhile, offer frontier performance at a "fraction of the cost" of competitors, Nadella claimed in the keynote. 

However, one of the most poignant examples of Microsoft tracking the shift towards cost efficiency is Agent 365 Business Analytics, a tool that aims to track whether enterprise AI investments are actually leading to returns. Rather than just tracking tokens spent, Business Analytics connects those tokens with work that's actually been completed, calculating returns by comparing the time that a human would take to complete those tasks with the amount that it cost the model. It also tracks things like success rate and user satisfaction with AI. 

Measuring ROI is "top of mind of every company right now," Charles Lamanna, executive vice president of Copilot, Agents, and Platform at Microsoft, told The Deep View. 

"We have a way where we can automatically detect what we think the outcome is of AI usage," Lamanna added. "Basically, what are the business KPIs you already track, and how are they being moved as a result of AI? Spending a million dollars on tokens for a business unit, it could be good or it could be bad. It all depends on what that million dollars did."

Focusing on efficiency in its releases at Ignite is a continuation of an ongoing trend for Microsoft. Efficiency was also a major theme of its updates at its Build event in June, with its first reasoning model sitting at just 35 billion parameters, compared to the multi-trillion-parameter models that other frontier labs are releasing. It's a move that makes sense for Microsoft: The tech giant has long been keenly focused on the pain points of enterprises. And now more than ever, enterprises are feeling the pain of increasingly expensive token bills. The shift towards efficiency as enterprises seek to hit the sweet spot on the cost to performance ratio, however, leaves a few questions up in the air. Are frontier models becoming commoditized? And if so, how does the drive towards cheaper AI, especially with the rise of ultra-low cost decision models like Jev and Microsoft-Decision-1.1, shift the economics of an industry with trillions of dollars hanging in the balance? 

Nat Rubio-Licht

TOGETHER WITH SIMA.AI

Want to Bring Intelligence to Your Product?

From drones and robots to smart cameras, more and more products are expected to see, decide and act on their own. The catch: real devices come with tight power budgets, real-time performance demands and no guarantee of a cloud connection.

That’s the problem SiMa.ai’s Modalix MLSoC and agentic Palette software suite solve. The new Palette 3.0 release brings several practical upgrades. Developers can now easily execute multiple physical AI models concurrently on a single chip, tune new camera sensors incredibly fast without recompiling or new code, fit 20B-parameter models on a 16GB module, and get up to 3.5x more tokens/sec with speculative decoding.

PRODUCTS

Mistral sees a market the AI giants may miss

Mistral has always targeted enterprises and technical use cases. Now, it's doubling down.

On Tuesday, the French AI lab announced it is building new models aimed at industrial engineering and other regulated sectors. The company said its goal is to give these fields access to frontier, sovereign AI models that offer the benefits of open models, such as full control over the model and the ability to own and customize it. It is focusing specifically on industrial engineering, which it calls AI’s second wave after software engineering, given how much the industry stands to gain.

"Engineering, similar to other regulated industries like finance and energy, must run on proprietary data, specialized models and systems no one else can switch off," said Arthur Mensch, CEO and co-founder of Mistral, in the post. "That is what sovereign AI delivers, and it is why the most demanding organizations in the world are building with Mistral.” 

The new models and capabilities build on Mistral Large 4, nicknamed Le Chonk, which the company claimed at launch was as capable as the best closed models while competing with open-weight models three times its size. That release focused heavily on coding and cyber defense, with Mistral calling it the world’s leading open model for agentic coding and state-of-the-art among open models for cyber defense. Another selling point is that it was trained in Mistral’s own European data centers, using less money and compute than closed labs typically need. This is particularly notable as the debate around data centers continues to pick up in the US.

Mistral shares that the industrial engineering models are designed to accelerate R&D and create products across a range of industries, including semiconductors, manufacturing, automotive, aerospace, and energy. The company's acquisition of Emmi AI earlier this year, which offers expertise in the physical AI space, will contribute to these efforts. No exact timeline was shared.

Open models keep gaining momentum, driven in large part by the steady stream of highly capable releases coming out of China. For enterprises, the appeal is clear: the biggest advantage is owning your models outright, customizing them to your exact business needs, and never having to worry that a provider will sunset them without warning. Closed models have real strengths too, particularly in reliability and security, since the provider typically handles maintenance. But the more companies I talk to, the more I see how widely their needs vary, and in that environment, more choice is better. I think there’s a sizable market for Mistral to capture, even if it won’t match the scale of what Anthropic and OpenAI have built.

LINKS

  • Gemini Notebook: turn your sources into interactive, shareable websites

  • iOS 27.3 beta: Apple Adds Daily Limits to Advanced Siri AI

  • Howie: The email and calendar secretary is joining Google Labs

  • Suno Visuals: Brings music to life with AI-video

  • Gemini: Google is expanding Tasks in Gemini using the Hydrogen 4 model

  • Koch: AI Solution Architect

  • Salesforce: Applied AI Architect (Principal/Lead)

  • EY: Agentic AI Engineer - Manager

  • Deloitte: AI Security Engineer Manager

GAMES

Which image is real?

Login or Subscribe to participate in polls.

POLL RESULTS

Do you think AI labs need to be more transparent about their third-party auditing?

Yes (85%)
Somewhat (8%)
No (7%)

The Deep View is written by Nat Rubio-Licht, Sabrina Ortiz, Jason Hiner, Faris Kojok and The Deep View crew. Please reply with any feedback.

Thanks for reading today’s edition of The Deep View! We’ll see you in the next one.

“Empty seats was the tell tale sign of which image was real. ”


“AI would never be able to resist showcasing it's brilliance by showing blurred faces.”


“The face expressions were clear and more realistic.”


“[This image] shows varying emotional states and different riders in and out of focus as would be expected. ”

“The people in [this image] didn't seem genuine.”


“Two people per cart. [This image] appears to have riders not in individual carts.”


“Repeat figures wearing white Ts in [this image]”


“The lighting in [this image] appears artificial.”

If you want to get in front of an audience of 750,000+ developers, business leaders and tech enthusiasts, get in touch with us here.