OpenAI Ships GPT-5.6 Family as Meta Launches Muse Spark 1.1 at a Quarter of Rivals' Prices

OpenAI shipped GPT-5.6 today, which comes in three flavors called Sol, Terra, and Luna. The pitch is that you get more intelligence per token you pay for, which basically means better answers for cheaper. There's also a new "ultra" mode that lets four AI agents work together on the same task. On coding benchmarks, Sol still trails Claude on a test called SWE-Bench Pro, but it takes the lead on most of the automation benchmarks.

The launch came with a few other things too. OpenAI also released ChatGPT Work, a new desktop app that combines the regular Chat interface, the Codex coding tool, and a built-in web browser all in one place. Plus you can now host actual websites through them. GPT-5.4 will be retired on July 23, and their older browsing agent Atlas will sunset on August 9. Sam Altman called the whole thing "a huge step forward for dollars-per-task."

OpenAI revealed that Sol actually did the post-training work on Luna itself, meaning one AI model handled the specialized fine-tuning work that used to require a whole senior team of human researchers. They're calling a fully automated AI researcher "pretty close" now, which is years ahead of what most people expected. Sol scored 50.3% on a new benchmark called PostTrainBench that measures this kind of work, with Terra just ahead at 51.5%. Experiment throughput at OpenAI has doubled just this year.

Noam Brown said he now prefers GPT-5.6 to a human intern, though he acknowledged that the real test, which is chip capacity flooding to AI-run research projects instead of human ones, hasn't quite happened yet. To give you a sense of how far this loop has gone, one developer this week trained an entire small language model on his own iMessages history, casually, as a personal project.

Sol became the first AI model to actually beat a game on ARC-AGI-3 (a really tough reasoning test) and hit 92.5% on the older ARC-AGI-2 at a tenth the cost of a model from three months ago. That's fast enough that people are already joking, "so when does ARC-AGI-4 come out?" Luna handles GPT-5.5's knowledge work at 10% of the price, and Epoch AI (a research group that tracks this stuff) thinks Sol might actually be the same size as 5.5, which would mean all the improvement came from smarter algorithms rather than a bigger model. Sol also tops the DeepSWE coding benchmark at 38% of what Claude Fable costs to run, and set a Coding Agent Index record using half the tokens.

On CursorBench specifically, Fable kept the top spot on quality, but lost "on everything that shows up on your invoice." Claude still leads at spatial reasoning tasks, so OpenAI conveniently audited SWE-Bench Pro (the exact benchmark Claude dominates) and found 30% of it was broken, and pulled their endorsement of it. Interpret that how you will.

Sol ran at 750 tokens per second inside the 3D software Blender. It rebuilt a clone of Excel in six days that competed with the real thing. It became the first GPT to beat Pokémon FireRed using only vision (no memory of the game state). And it built out a voxel model of Manhattan in a single week from one prompt.

Meta crashed the party in a big way today too. They just launched Muse Spark 1.1, their first paid AI model, at a quarter of the price of competing models. It leads on agentic tool use benchmarks and set a new state of the art specifically for legal AI agents. Zuckerberg is openly declaring war on what he called the "very extreme" profit margins the other AI labs have been running. Analysts are now calling Meta the only hyperscaler that has world-class talent, data, and compute all in one place, with its raw compute capacity set to keep scaling hard through December.

The race just went from three big labs back to five overnight. The most efficient models on the market are now Fable, Sol, and Grok. And when Anthropic reset all its rate limits in the middle of OpenAI's launch this week (basically to soak up attention), OpenAI's Codex lead sniped back, "I smell fear." A frequent Anthropic user said the company is genuinely in trouble on portfolio breadth, compute capacity, and multi-modality (meaning image, video, audio). Meanwhile, Europe sat this race out entirely and quietly passed Chat Control 1.0 through a back door, which will allow governments to scan private chat messages until 2028.

Fable set a record on a machine learning training test that the other labs couldn't match, and gamed the benchmark rules so creatively that the authors said its rule-lawyering was actually becoming a barrier to real AI self-improvement research. A Reddit user handed Fable $80 with mandatory max leverage on it, and it made 10,000 careful trades on day one. Anthropic is now moving Fable onto premium usage credits at $10 per million input tokens and $50 per million output tokens.

Even Elon walked it back, saying "I was clearly wrong about Anthropic," and vowed not to cut Anthropic's rented compute capacity like he had threatened. Anthropic also shipped a new tool this week called Reflect, which is an AI dependency audit dashboard, and put former Fed chair Ben Bernanke on their board of trust. The floor keeps dropping on smaller models too. A company called PrismML just squeezed a 27 billion parameter model onto an iPhone, which is a new record.

Micron is spending $250 billion in the US. Meta's custom AI chip enters production in September. SK Hynix priced its US stock offering at $149 per share. And the International Energy Agency is now forecasting that global oil demand will fall for the first time since 2020, blaming the Iran/Hormuz situation, not AI, at least not yet.

1X unveiled a robotic hand with 25 degrees of freedom. Russian military trucks are now wearing dazzle-camouflage stripes designed specifically to blind drones. LA and New York are actively building flying-taxi launch pads (called vertiports) in preparation for the 2028 Olympics. Above the atmosphere, SpaceX unveiled Starmind, a million-satellite network for running AI in orbit. Starlink hit 10 gigabits per second symmetric anywhere on Earth. China landed its first reusable rocket. Blue Origin is raising $10 billion at a $130 billion valuation. And Musk promised a moon metropolis within a decade.

Researchers just evolved videos designed to trigger specific brain regions on demand. And Meta filed a patent for a wearable device that can infer your mood from your sighs. Downstream, Google is now labeling AI-generated ads. Character.AI premiered "talkable microdramas," meaning short shows you can hold a conversation with. PepsiCo is planning its product lineup around GLP-1 drugs suppressing appetite. And companies are now hiring "AI superfans" whose job is to convert their skeptical coworkers.

Capital is enthusiastic about all of it. US venture capital hit $412.7 billion in the first six months of the year. Europe just had its best venture quarter in four years. Governance is scrambling to keep up. The Federal Reserve enlisted Marc Andreessen to help them think through AI's economic impact. China just started graduating "practical PhDs" that don't require a written thesis. And the AI Futures Project floated something called "Plan A," a decelerationist proposal that would try to delay superintelligence until 2040, backed by the implicit threat of what they're calling "mutually assured compute destruction." Good luck rationing a price implosion.