AI BreakfastOpenAI’s new GPT-5.6 features an ‘ultra mode’ to run parallel agents
Jul 10, 2026
In partnership with
Good morning. It’s Friday, July 10th.
Ai4 (America’s Largest AI Conference) is coming to Las Vegas August 4–6 at The Venetian!
Please let me know if you are attending! I’ll be there with a media pass. Would be great to meet some of you in person!
-Jeff
AI Breakfast
You read. We listen. Let us know what you think by replying to this email.
You've seen the AI demos. Viktor does it without you watching.
The AI tool you tried last quarter waited for a prompt, hallucinated a number, then asked if you'd like a summary.
Viktor opened a PR at 2am, rebased it against main, ran your test suite, and posted a note in #eng: "Two flaky tests in payments service, both pre-existing. Recommended merging after fixing them." Then drafted the customer reply for the support ticket the bug created.
That's 619K autonomous actions per day across 20,000+ teams. Not chat replies. Real work shipped to GitHub, Stripe, Linear, Notion, and 3,000+ other tools, from inside Slack and Microsoft Teams.
You don't supervise him any more than you supervise a senior engineer.
SOC 2 certified. Your data never trains models.
"It's what you probably originally thought AI was going to be when you first heard of it in sci-fi movies." Tyler, CEO.
OpenAI’s new GPT-5.6 features an ‘ultra mode’ to run parallel agents
OpenAI released the GPT-5.6 family in three tiers: flagship Sol, intermediate Terra, and budget-friendly Luna. Built to optimize token consumption, Sol improves agentic coding efficiency by 54% and sets a state-of-the-art mark of 80 on the Coding Agent Index, outperforming Fable 5 at one-third the cost.
Despite government security review delays, the release prioritizes defensive cybersecurity. Sol introduces ultra mode, coordinating four parallel agents via Programmatic Tool Calling to run lightweight code and cut cyber benchmark token usage by a third.
OpenAI also introduced ChatGPT Work across mobile and desktop to automate multi-step documentation, prompting the August 9 retirement of its standalone Atlas browser. API input pricing ranges from Luna at $1 to Sol at $5 per million tokens. OpenAI has confirmed GPT-5.6 is the preferred model for Microsoft 365 Copilot. To help businesses actually deploy all this stuff, OpenAI’s deployment arm just acquired the applied-AI consulting startup Northslope.
Finally, OpenAI is completely overhauling voice with GPT-Live. It uses a full-duplex architecture so you can interrupt it naturally, and it splits the pipeline, the voice engine keeps the chat moving smoothly while GPT-5.5 handles deep reasoning and web searches in the background.
All of this arrives amid a major internal shakeup: AGI lead Fidji Simo is shifting to an advisory role due to a chronic illness, leaving President Greg Brockman to take the helm of OpenAI's newly unified agent and product strategy.
Related videos:
This is the new ChatGPT Voice, powered by GPT-Live
Anthropic drops context costs for Fable 5; rolls out ‘reflect’ to track your AI habits
Anthropic is trying something new: checking in on your relationship with AI.
The company just dropped a beta ‘Reflection’ dashboard for all users with Memory turned on, designed to track your chat habits over the past year. It feels a bit like a screen-time report, mapping your behavior against Anthropic’s 4D AI Fluency Framework while nudging you with quiet hours and break reminders to keep your human thinking sharp. It keeps things private by locking out incognito chats, files, and health queries, and it's headed to Claude Cowork next.
Anthropic also reset all five-hour and weekly usage caps across the platform.
But the real structural change is how it's countering cheaper rivals like OpenAI’s Sol and xAI’s Grok 4.5. Instead of running expensive linear workloads on Fable 5, a new multi-agent framework lets developers use Claude Managed Agents to treat Fable as a manager that delegates execution to Sonnet 5. By isolating sub-agent caches to wipe out duplicate context costs, this routing keeps 92% of SWE-Bench Pro coding power at 63% of the price, and hits 96% of web automation benchmarks at less than half cost.
To monitor how these economics impact the labor market, Anthropic appointed former Federal Reserve Chair Ben Bernanke to its independent Long-Term Benefit Trust.
SpaceXAI undercuts frontier market with aggressive Grok 4.5 token pricing
SpaceXAI, fresh off its $60 billion acquisition of Anysphere, has launched Grok 4.5. Built on a 1.5-trillion parameter architecture and trained across tens of thousands of Nvidia GB300 GPUs, the model targets actual enterprise utility. Because developers seek work-tested data over influencer hype for Grok 4.5 evaluations, the release leverages asynchronous reinforcement learning over trillions of developer tokens to compete directly on speed and token compression.
VentureBeat reports that SpaceX's Grok 4.5 launches at half the price of rivals and could rattle Anthropic and OpenAI. This architectural framing yields an aggressive pricing structure:
Grok 4.5: $2 in / $6 out
GPT 5.6: $5 in / $30 out
Opus 4.8: $5 in / $25 out
Independent trials reveal that Grok 4.5 wins on speed and cost while Anthropic maintains the reasoning crown in latest coding test. While Claude Fable 5 natively executes complex 3D Rubik’s Cube simulations that stump rivals, Grok 4.5 counters with rapid 80 to 110 tokens per second outputs. It closes long-horizon workflow gaps by using 4.2 times fewer output tokens than Opus 4.8 on identical SWE-Bench Pro tasks, effectively balancing out minor deficit scores like its 53% on DeepSWE 1.1 versus Fable 5’s 70%.
The model has also proved its specialized training in specific verticals, securing the top spot on Harvey's Legal Agent Benchmark. Demonstrating its real-world execution speed, a developer trial showed Grok 4.5 builds functional FPS game in under an hour using agentic loops within the Grok Build harness, autonomously managing everything from asset scraping to localized execution loops
Frontier models and product moves
Google rolls out AlphaEvolve to let AI rewrite its own code (X)
Google trains SensorFM health model on a trillion minutes of raw wearable data (X)
Muse Spark 1.1 is Meta's new push into paid workplace AI (Meta)
Coding favorite Cursor leverages SpaceXAI to train its new Sand agent (The Information)
Zhipu AI stock spikes 11% on a massive $4 billion fundraising plan (Barron's)
US developer traffic to Chinese AI models spikes past 40% (CNBC)
DeepSeek V4 captures token share by undercutting proprietary API pricing (OpenRouter)
ByteDance introduces spatial layer breaking in Seedream 5.0 Pro UI design (ByteDance)
A single camera is all Mistral needs to guide robots through unknown spaces (The Decoder)
Google Gemma 4 gains local hardware acceleration on React Native apps (X)
Agents and the agentic stack
Viral AI agent OpenClaw gets a governance foundation (TechStrong)
New Perplexity advisor tool stops systems from calling expensive frontier models (Decrypt)
Cognition debuts SWE-1.7 with a self-compacting memory to survive six-hour coding tasks (X)
Hermes introduces a three-tier progressive disclosure system to save context tokens (X)
New Lingbot-World-2 architecture solves environmental drift in real-time simulations (X)
A simple statistical calculation reveals the absolute limit of your multi-LLM setup (VentureBeat)
New open-source toolkit teaches AI agents how to code like Apple (X)
A new MCP connector links ChatGPT and Claude directly to team documents (SiliconAngle)
Former GitHub chief builds a Git network to store AI agent prompts (SiliconAngle)
Berkeley and Princeton deploy open science autoresearch tracks on FrontierCS (X)
Business, labor, and institutions
Tencent prepares a $2B buyback of Manus after Beijing blocks Meta ownership (Reuters)
Lovable reportedly in talks to double its valuation to $13.2B (TechCrunch)
Character.AI launches microdramas where viewers can text the characters (TechCrunch)
Amazon documents reveal a costly new Alexa AI project (Business Insider)
Black market demand grows for copper and infrastructure bound for AI data centers (Vice)
Standard Bots targets the impossible tasks of manufacturing (Automate)
Research and benchmarks
Apple meets with Caltech spinout that compressed a 27B parameter AI model onto an iPhone (MacRumors)
Project Triad by the US government creates a new pipeline for quantum machine learning data (NSF)
Watch: New 25 degree of freedom hands put the 1X NEO humanoid ahead in tactile precision (X)
Timbal is an all-in-one enterprise AI platform that lets you build, deploy, and scale production-ready agents, deterministic workflows, and custom interfaces.
ExploreYC is an open-source REST API delivering structured JSON data on over 6,600 Y Combinator and a16z companies.
Talkory runs your prompt through five frontier LLMs simultaneously, using recursive error-correction to deliver unified, consensus-driven answers.
AnySearch v2.1.0 is a private, agent-native search infrastructure delivering fast, structured, multi-domain data via API and MCP.
Toyo is an AI executive assistant managing your schedule, meeting prep, and email triage directly through iMessage.
Thank you for reading today’s edition.
Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.
Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!
Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.
Keep Reading
AI Breakfast
Curated weekly analysis of the latest AI projects, products, and news