In partnership with
Reading time: 6 minutes
🗞️ In this edition
OpenAI Cut The Floor In Half
Sponsored: Aligned
Ema Raised $77 Million On A Prediction
Z.ai Was Taking The Whole Repo
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
Snorkel Tripled On Training Data
Go.AI Raised To Keep Data Home
Identity Digital Spun Out Agent ID
4 must-try AI tools
OpenAI halved its API prices on Tuesday and Anthropic answered the same day. Ema raised $77 million on a claim about the software you already pay for. And a researcher worked out where Z.ai's coding tool had been sending your repository.
Keep reading.
Source: Bankinfoseccurity.com
Claude Opus 5.5 landed the same day.
WHAT HAPPENED
GPT-6 Sol costs $2 per million input tokens and $10 out, GPT-6 Luna costs $0.10 and $0.50, and both are half what their GPT-5.6 equivalents charged. OpenAI says Sol takes 56.4% on Agents' Last Exam at 60% below Claude Opus 5's cost and 60.5% on OSWorld 2.0 at 80% below, and makes about half as many factual mistakes as the model before it. Anthropic released Claude Opus 5.5 the same day at $4 in and $20 out, 20% under Opus 5 and 60% under Fable 5.1, with a one million token context window. Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1, and cites a 200,000 line codebase audit finishing in under three hours where Opus 5 needed more than twenty. Neither lab has published a same-harness comparison of the two.
WHY IT MATTERS
A cut this size usually follows a cheaper way to serve the model. Two labs arriving at the same cut on the same day is not two independent efficiency wins, it is both of them reading the same competitor off the same spreadsheet. What that does downstream is quietly close a gap a lot of companies were built inside. A year of startups raised money on the margin between what a frontier model costs to call and what a narrow fine tuned model costs to run, and that margin lost half its width in one afternoon. The floor is now set by whoever is willing to lose the most on inference.
OUR TAKE
The obvious read is a price war that buyers win. We think it is a land grab with a deadline, and the tell is that both labs cut the cheap tier hardest. Luna at $0.10 in is not really aimed at Anthropic. It is aimed at the moment you consider running an open model yourself, and both labs want that question settled before open weights make the answer obvious.
Watch whether either price holds. If GPT-6 Luna still costs $0.10 per million input tokens in June 2027, this was a real cost reduction passed on to you. If it has drifted up, or the cheap tier has quietly been rate limited into a trial, it was customer acquisition priced as a product.
From Our Sponsor:
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
Creaegis led it, with Accel, Section 32 and Prosus all raising their stakes, taking the company to $140 million raised since 2023 at roughly four times its 2024 valuation. Ema says it has booked more than $150 million in multiyear contract value across over 50 enterprise deals, with more than a million active enterprise users, gross margins near 80% and net dollar retention around 180%, on a headcount of about 200. Customers include NTT DATA, Hitachi, ADP, PwC, Google, KPMG, Wipro and Microsoft.
How did today's stories land?
Source: Adsoftheworld.com
Ferstar found that a feature called Repo Wiki packaged whole user workspaces, complete project histories included, and sent them to Alibaba Cloud after building documentation pages, with no setting to switch it off. The uploads were encrypted under a private key held only on Z.ai's servers, so users could neither read nor delete their own code. Z.ai apologised on social media, said the data never trained a model, cited outside assessments confirming deletion, removed Repo Wiki and open sourced ZCode on GitHub.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
🧮 Nokia open sourced AnyJev under Apache-2.0, a training free layer that turns any open model into a calibrated decision model. On BANKING77 it cut answer flips from 23% to 7.3%. → Grab the library
🎙️ Kyutai released Voice of Reason, two 9B open weight checkpoints that solve spoken maths with no transcription step at all. GSM8K went from 27.3% to 77.1%, on one H100. → Take it for a spin
🧵 Tsinghua researchers got models talking through each other's KV caches instead of through text. Accuracy up 9.6 to 11.9 points, and one handoff took 90 milliseconds where text took 1,312. Apache-2.0 on GitHub. → Read the method
🎧 Spotify opened its recommendation engine to plain language instructions. US Premium users 18 and up get it from today, in beta, and the Home feed shifts within a few hours. → Tune your feed
Snorkel Tripled On Training Data – The $350 million Series E at a $3.5 billion valuation was led by Insight and S32, on a data-as-a-service line that grew 18 times and crossed a $375 million run rate.
Go.AI Raised To Keep Data Home – Updata Partners led an $85 million Series A for on-premises AI hardware and software aimed at banks, defence and healthcare, priced at a fixed fee rather than by usage.
Identity Digital Spun Out Agent ID – Known Systems AI is pushing DNSid, a birth certificate for AI agents built on DNS, public key infrastructure and a blockchain ledger, with an Internet-Draft now at the IETF.
Taskade: Turns the plan into tasks, then works through them alongside you.
Jasper AI: Writes the campaign in your brand voice instead of a generic one.
Midjourney: Makes the image you were about to describe to a designer.
MurfAI: Gives your deck a voiceover that does not sound like a robot.
Three stories about the part of the deal you do not get to see. OpenAI and Anthropic set the price of reasoning in public and keep the cost of serving it private, so nobody outside can tell a real efficiency gain from a subsidy. Ema is selling the argument that your software licence was the expensive part all along. And Z.ai held the only key to code its users did not know it had taken. The invoice is the visible bit. It is rarely the part that decides anything.
Look behind the price.
Interested in featuring your services with us? Email us at [email protected]
