The new ChatGPT desktop app puts all of OpenAI’s eggs in the same basket. (Picture: OpenAI)UPDATE: OpenAI seems to have removed the chat history and put the chat itself in a popup window. If you don’t like this in the new app, and a lot of people don’t, it leaves the ChatGPT Classic app untouched, which can be used for pure chats like the old version. It will still recieve updates and won’t become obsolete.
—
Along with the release of GPT-5.6 today comes a new app that looks unremarkable at first, but is a one-stop-shop for all of OpenAI’s features.
The latest addition is called ChatGPT Work, which is billed as an agent that gathers information from your apps and workflows and turns them into spreadsheets, presentations, docs or PDFs and just about anything you want — even web apps.
It’s getting integrated into a much more interesting, upgraded ChatGPT desktop app that will combine Codex, Work, agentic browsing and the normal chat interface.
GPT-5.6-Luna is competitive with recently launched cheap models, while Sol is more expensive. (Picture: OpenAI)GPT-5.6 was, as promised, released widely to all of OpenAI’s customers today — and will take around 24 hours to fully propagate.
The model reaches state-of-the-art status through many benchmarks, beats Fable and Mythos often, and scores a record on Arc-Agi-2 of 92.5% while being the first to actually solve a puzzle on Arc-Agi 3, measuring how it handles completely unknown situations.
GPT-5.6 comes in three versions; Sol as the flagship, top rated model, Terra as the mid-tier GPT-5.5-level model at half the price, and Luna as the dirt cheap model, likely comparable with GPT-5.4.
The pricing for Sol is at $5 input/$30 output, Terra is at $2.50 input/$25 output, and Luna ticks in at $1 input and $6 output, placing it in a good position to compete with the recently released Grok 4.5 and Muse Spark 1.1.
The Free and Go tiers on ChatGPT only get access to the Terra model, while Plus, Pro, Business and Enterprise can choose between all three, and set effort levels.
Pro and Enterprise users get Sol Pro in addition, which provides «the highest quality results on complex tasks.»
GPT-Live is better at waiting for its turn to speak, and you can interrupt for a better back-and-forth. (Picture: OpenAI)150 million people use their voice to interact with ChatGPT per week, OpenAI says, and many have waited patiently since 2024 for an upgrade.
The new GPT-Live uses a «full duplex architecture» that separates out the underlying language model and continuously processes speech.
This means it can do «mmm’s» and «yeah’s» during a conversation to indicate it is actively listening, but more importantly it can listen and speak at the same time.
OpenAI says GPT-Live makes decisions «many times per second» on whether to speak, listen, pause or interrupt, letting it engage more naturally in conversations — and do live translation.
The new model delegates more complex tasks to «the latest frontier model» (currently GPT-5.5) for reasoning or web search and gets back to you as soon as it is ready, sometimes showing cue cards for at-a-glance responses in the app.
GPT-Live is available today by tapping the Voice-button in the app or on the web for Go, Plus and Pro subscribers, and there is a GPT-Live-1 mini for the free tiers.
OpenAI has announced the general release of GPT-5.6 tomorrow. (Picture: OpenAI)GPT-5.6 was officially unveiled a little over a week ago, but got hampered by a US government intervention for cyber risks and national security, having been shown in benchmarks to beat Anthropic’s Mythos.
At launch, it got released only to some 20 «trusted partners,» and OpenAI promised a general release «in the coming weeks.»
Now those weeks have passed, the government has lifted its restrictions without further explanation, and both OpenAI’s X account and CEO Sam Altman are posting that it will «launch publicly this Thursday.»
This comes after extensive testing by the Center for AI Standards and Innovation within the Department of Commerce, aided by experts from OpenAI, who stayed on call for «potential questions,» Axios reports.
No account has been given as to what specifically caused the delayed launch or what conditions the government had for its release.
The Codex keypad looks like an iteration of an existing product, not something entirely new. (Picture: OpenAI)The OpenAI Developers account on X teased a video of a square device with backlit keys made in collaboration with Work Louder, along with the date July 15th.
It’s likely a small keypad for common Codex shortcuts, as mentioned by the post, and it will be the first hardware device (co-)developed by OpenAI. Not a flashy new phone or the highly anticipated AI hardware device, which would take years to develop.
Work Louder is mostly famous for mechanical keyboards and custom keypads, and makes a $144 13-key wireless pad, complete with a tiny joystick.
The device being teased in the video looks like a carbon copy of the Creator Micro 2 pad described above, save for the backlit keys.
That should offer a hint for how OpenAI will be playing its hardware — as a premium Codex add-on, likely also with a premium price.
The flagship GPT-5.6 Sol is the new state of the art for cyber capabilities. (Picture: Adobe)Caught in yet another government AI debacle, the new models won’t be publicly released until «the coming weeks.» It is getting presented today, and released to «a small group of trusted parties,» said by Axios to number in the twenties.
— We don’t believe this kind of government access process should become the long-term default, writes OpenAI, and says — we [are working] with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.
In its presentation, OpenAI details three models in the GPT-5.6 series. Sol is the state-of-the-art Mythos-beating model, while Terra is «a balanced model for everyday work,» and delivers on par with GPT-5.5 at half the cost, and Luna is the cost-focused model «delivering stronger capability at our lowest cost.»
Strong on cyber
The models, particularly Sol, are very strong on cyber defense, and are able to identify security vulnerabilities, but not able to execute automated attacks due to strong safeguards, built using 700,000 A100-equivalent GPU hours of red-teaming.
Sol beats Mythos 5 on coding and cybersecurity and uses fewer tokens for better performance on biology tasks. It also comes with a new max mode, which gives more time for better reasoning, and an ultra mode which uses subagents for more complex work.
Prices for Sol are at $5 input/$30 output, for Terra is $2.50 input/$15 output, and Luna is at $1 input/$6 output, and general availability is up to higher powers.
GPT-5.6 is officially under government review due to «security concerns.» (Picture: generated)First reported by The Information, OpenAI has agreed to the Trump administration’s demands that it stagger the release of the upcoming GPT-5.6 over security concerns.
The model is said to be on par with Anthropic’s Mythos and Fable 5, which were abruptly pulled from the market after an export ban from the administration just two weeks ago.
Axios is quoting sources familiar saying that GPT-5.6 will only be released to a «small set of government-approved partners,» as the White House previews its abilities.
This preview period is expected to last «a couple of weeks,» CEO Sam Altman told staff in an internal OpenAI meeting, TechCrunch reports.
During this time, the administration itself will supposedly approve access to the model on a case-by-case basis, according to The Verge.
The Trump admin issued an executive order on AI about three weeks ago, setting up a voluntary mechanism for AI labs to submit their models for 30 days of pre-release testing. This gave the government 60 days to come up with criteria, but it isn’t quite ready yet — hence this ad-hoc approach.
The Jalapeño chip offers substantially better performance per watt than anything on the market. (Picture: OpenAI)The new inference chip, developed with help from OpenAI’s models in collaboration with Broadcom, has compressed what is normally a glacial, years-long process into just a short sprint.
— We believe [this] to be the fastest ASIC development cycle ever achieved in high-performance advanced semiconductors, OpenAI says.
The chip is designed specifically for OpenAI’s needs on current and upcoming language models, and to combine the power of current AI accelerators with the shorter latency of «specialized systems,» like Cerebras’ hardware.
— The world is moving to a compute-powered economy, says Greg Brockman, President and Co-Founder of OpenAI. — Jalapeño is part of our long-term full-stack infrastructure strategy to make compute more abundant.
The new silicon will be deployed at gigawatt-scale data centers just as fast as it was made, starting with Microsoft «and other partners» already in 2026, Broadcom says. It is only the first design in what the companies expect to be generations of products.
Jalapeño is already running LLM workloads in the lab, delivering «substantially better» performance per watt than «the current state-of-the-art,» OpenAI says.
The new Cyber model not only finds flaws, but offers to fix them automatically. (Picture: Adobe)Less fabled than the recent Mythos releases, GPT-5.5-Cyber is just as capable, OpenAI claims, and has benchmarks to show it.
The update scores 85.6% in CyberGym versus 83.8% for Mythos 5, and is likewise only available to vetted cybersecurity professionals.
— AI has changed the physics of cybersecurity, OpenAI says, — The bottleneck historically has been finding vulnerabilities, but now defenders are overwhelmed with the number of vulnerabilities found.
Therefore, their Daybreak suite, which includes the Cyber model and Codex Security, moves beyond just threat scanning to actually offering vulnerability patches to fix them.
Since the launch as preview in March, OpenAI claims to have scanned over 30 million commits across 30,000 codebases, with humans having marked 70,000 findings as fixed and 500,000 findings having been fixed automatically.
OpenAI made $2 billion a month by the end of 2025. (Picture: Adobe)The spending came on a revenue of $13 billion, with a one-time accounting charge of $30 billion, The Financial Times reports, citing people familiar with the audited figures.
This charge was related to a change in governance from a non-profit to a Public Benefit Corporation. Previously, the FT explains, investors were paid interest on their capital, and this was accounted for as liabilities. As a PBC, they now pay in equity, and the cost won’t be reoccurring.
The numbers also show that the company spent $19 billion on developing models and research, and around $6 billion on sales and marketing.
At the end of 2025, OpenAI was earning $2 billion in revenue per month, as opposed to $1 billion per quarter in 2024, which the FT says makes it one of the fastest growing companies in history.
Without the one-time accounting charge, OpenAI’s losses for 2025 were «only» $8 billion.
OpenAI is estimated to be worth well over a trillion dollars on the open market. (Picture: Shutterstock)OpenAI has put up a short notice on its website informing of the filing, saying that they haven’t decided on the timing yet:
— It may be a while because there are things we want to do that are likely easier as a private company, the message says, and adds: — But it’s a complicated set of tradeoffs and this gives us the option to go public sooner if that ends up being best.
This is just the first step of eventually filing publicly and paving the way for a proper market entry, which should give journalists and financial analysts a better look at their books, including risks and opportunities and real revenue numbers.
OpenAI’s most recent fundraise in April valued the company at $852 billion and saw companies such as Amazon, Nvidia and SoftBank pour in tens of billions each for a round worth $110 billion.
They said at the time that they had surpassed 900 million weekly users and had 50 million paying subscribers, and 6x the monthly web visits than the next AI app.
Anthropic filed confidentially just last week, and along with SpaceX, it sets up a season of some of the largest market entries ever. All three companies are expected to reach valuations of a trillion dollars or more.
OpenAI’s superapp isn’t just a convenience, it should also be a revenue driver. (Picture: Shutterstock)Much has been said about OpenAI’s coming «superapp,» combining Codex with ChatGPT and now also agents, but the Financial Times is now putting a date certain on it.
The new app is not just a convenience for users who use chat and coding and agents to have it in one place, but it is also a strategic play by OpenAI to funnel them into more lucrative businesses areas — particularly when it comes to agents.
Thibault Sottiaux, who leads OpenAI’s core product and platform efforts, tells the FT that the app will give you a personal agent to do stuff for you, be it personal or work, doing far more than the classic Codex, which now performs computer use and does office tasks.
In fact, 20% of Codex users are now knowledge workers, OpenAI reported last week. This category is also growing three times as fast as other use cases.
The concept is that you will have access to the new app from everywhere you go, from mobile to desktop to the web, so you could even talk to it in your car.
The idea behind the app is to boost revenue and ease enterprise adoption, now counting 2 million Codex users and accounting for 40% of revenue, Reuters reports.
The same goes for remote control with ChatGPT on iOS or Android, and you can also use a Mac to control Codex on Windows.
Codex computer use opens PDFs, spreadsheets, slides and docs and can even handle SSH connections. It can control «any app on your computer,» and can work in the background so it won’t bother your workflow.
The remote control feature can answer questions, review code, handle approvals or let you change directions on Codex right from ChatGPT on your phone.
You can download the app here, and it is available for Plus, Pro, Business, Edu and Enterprise plans.
OpenAI is still leading the pack, but is facing increasing headwinds. (Picture: Adobe)The Information (paywalled) says sources are telling them that OpenAI landed a billion dollars ahead of Anthropic last quarter, but «only» has an annualized run rate of $30 billion versus Anthropic’s $45B.
The AI lab estimates that its adjusted income margin was at -122%, meaning it lost $1.22 for every dollar earned, driven by high investment costs for infrastructure and compute.
The growth was driven mainly by Codex and enterprise customers, the report said — although ChatGPT growth stalled at about 905 million weekly users in the latest quarter.
OpenAI is said to reveal its IPO plans «within weeks,» Reuters reports, and possibly as soon as «this Friday,» according to CNBC.
Its most recent funding round valued the company at $852 billion, just short of Anthropic’s $900 billion — but their funding round came a little later.
OpenAI is off the hook for their for-profit organization. (Picture: Adobe)The jury spent less than two hours Monday to deliberate on the case, and dismissed Musk’s claims for being filed after the statutory limitation of three years.
The trial had been brought in 2024 over a 2019 decision by OpenAI to create a for-profit arm, which was later expanded into a public benefit corporation — allowing OpenAI to raise massive funds and possibly go public at a later stage.
Elon Musk was a co-founder of OpenAI in 2015, but left the board in 2018. He argued in court that it was set up as a charity, and his lawyers claimed that OpenAI’s executives had enriched themselves at the cost of the charity’s mission.
As the court decided on a technical issue, the jury never touched the actual merits of the case, Musk argues, and vows to appeal.
Announced in April, 
