The 48-hour AI arms race: ByteDance, Meta, and OpenAI all moved at once
Photo: N43 and HermesIn a 48-hour window, ByteDance revealed a 10 trillion parameter model in training, Meta shipped a terminal coding agent aimed at Claude Code, and OpenAI made its flagship free for a billion users while quietly confirming a larger successor at release candidate status.
01ByteDance goes for frontier scale
The Financial Times reported Friday, citing people with knowledge of the matter, that ByteDance is training a model with as many as 10 trillion parameters. Reuters picked up the story directly. The scale is not a marginal step over what Chinese labs have been producing; it is a different category entirely.
For context: Moonshot AI's Kimi K3, previously the headline number from China, sits at 2.8 trillion parameters. Before K3 shipped, the domestic leaders were Moondrop's Longat 2.0 and DeepSeek's V4 Pro, both at 1.6 trillion. The jump from where that ecosystem stood six months ago to where ByteDance is aiming is close to an order of magnitude.
Reported and estimated parameter counts for leading AI models. US frontier model sizes are industry estimates; ByteDance is in pre-training and not yet released. Source: Financial Times, Reuters, industry estimates.
The honest caveat: parameter counts get waved around as though they settle arguments. They do not. Parameters are the numerical settings a model learns from data to recognize patterns and generate answers. They are a rough measure of scale, and scale correlates with capability, but the two are not the same thing. Plenty of bloated models have lost to smaller, better-trained ones.
What makes the report serious is the comparison the FT drew. Direct benchmarking against US models is difficult because Anthropic and OpenAI do not disclose parameter counts for Fable, Mythos, or GPT-5.5. Nobody publishes those numbers anymore. What the FT has instead are industry estimates, and those estimates put Anthropic's most advanced system, Mythos 5, at roughly 8 trillion parameters, with Fable 5 around 5 trillion. That means ByteDance is not building something in the neighborhood of the Chinese frontier. They are building something in the neighborhood of the global one, potentially larger than the biggest model Anthropic is known to have.
02Why 10 trillion parameters changes the game
Reuters noted directly that they could not immediately verify the report, and ByteDance did not respond to a request for comment. There is also a timeline element: the model is in pre-training, which typically runs three to six months before fine-tuning and release. Nothing ships tomorrow. But the direction of travel is unmistakable. Chinese firms keep accelerating release cycles, and the whole balancing act they are running is building more powerful systems without making them prohibitively expensive to serve. At 10 trillion parameters, that second half gets a lot harder, which tells you how badly they want the first half.
03Meta enters the coding agent war
Meta shipped the beta of Muse Code, an AI coding agent that runs in your terminal, powered by their latest model, Muse Spark 1.2. Zuckerberg posted about it personally, saying it can complete complex software engineering tasks in large code repositories: analyzing projects, planning modifications, writing code, running tools, and verifying results. That is the whole loop. It means Meta, OpenAI, and Anthropic are now standing on the exact same ground, competing for the exact same developers.
It is macOS and Linux for now, installed with a single terminal command. The pitch is that you hand it a full requirement rather than a snippet: fix a bug spanning multiple modules, add a feature, refactor a large project. It reads the codebase first, builds a plan, modifies files, runs tests, and keeps adjusting based on what comes back.
Architecturally, the core is a fairly simple agent loop, but Meta paired it with a set of asynchronous background agents that boost the main one. The distinction that matters: those background agents run continuously across the entire session instead of being spun up fresh for each task. Because they persist, they are not recollecting the same context over and over. They can independently carry out follow-up steps and decide for themselves when results are worth feeding back to the main agent. The payoff is lower latency and a main agent that leans on you less during complex multi-step work.
The runtime design is the smartest piece. Muse Code keeps a local event log where model calls, tool runs, approval operations, and code modifications are all recorded in sequence. That log is the only trusted source of data, which means the entire run can be replayed accurately and safely restored after a restart. If the program crashes, the agent resumes from the exact point of interruption rather than starting over or drifting somewhere weird. For long-running tasks, that is the difference between a tool you trust overnight and one you have to babysit.
Built-in skills round it out: /plan decomposes a task into a plan that requires your approval before anything executes. /grill repeatedly stress-tests that plan until the solution is reliable enough. /goal keeps pushing toward whatever target you specify until the task is actually complete.
Muse Spark 1.2 benchmark results: second everywhere, first nowhere. Only Opus 5 Max consistently beats it. Sources: TerminalBench, DeepSuite, Meta internal evaluation.
04Benchmarks: second everywhere, first nowhere
The benchmark picture is consistent and frankly a little pointed. On TerminalBench 2.1, which evaluates how well agents complete tasks in a terminal environment, Muse Spark 1.2 is beaten by exactly one model: Opus 5 Max. On DeepSuite 1.1, it comes in behind Opus 5 Max and GPT 5.6 Terramax. That benchmark is no joke: 113 tasks across 91 repositories in five languages (TypeScript, Go, Python, JavaScript, and Rust), with every task shipping a manually written functional verification program plus regression tests. On Meta's internal coding benchmark built from 440 tasks derived from real internal pull requests, it is once again only Opus 5 Max ahead.
Second everywhere, first nowhere. Strong debut, and it tells you precisely who Meta was benchmarking against in the mirror.
Meta attributes the gains to three things. First, co-training: Muse Spark 1.2 and Muse Code were trained together so the pair performs in tandem, with agent operation trajectories introduced through rejection sampling training, target execution context compression, subagent links, and the Muse Code tool set folded into training. Second, long-horizon capability: large-scale training on full repository generation, big end-to-end projects, and automated research where the model sequences work through planning, holds direction via goal conditions, and uses context compression. Third, self-improvement: Meta used Muse Spark 1.1 to generate highly difficult programming environments and instruction-following templates, then had the model evaluate how well candidate solutions met requirements, building scalable training data for 1.2.
Muse Spark 1.2 API pricing. The contributor tier is not a price point; it is a recruitment drive. Source: Meta.
05The contributor price is a recruitment drive
Pricing is where they get aggressive. Standard runs at $125 per million input tokens, $15 for cached input, and $4.25 per million output tokens. But the contributor version is $0.10 per million input, $0.002 cached, and $0.20 output. That is not a price point. That is a recruitment drive.
06OpenAI makes its flagship free for a billion users
GPT 5.6 Luna is now completely free with unlimited text conversations for roughly a billion users worldwide. Luna is the smallest model in the GPT 5.6 family, built for ultra-fast response and low cost, sitting under Terra and Saul in the lineup. It becomes the default for both free and Go tiers, directly replacing GPT 5.5.
They also added a think button to free ChatGPT. When you hit something difficult, you can make the model reason longer before answering. This is the first time free users have been able to actively turn up reasoning intensity themselves.
The fine print matters: unlimited applies to text conversation only. File uploads, image generation, voice, and everything else still run on original quotas, and you will still hit caps.
07Saul stops rambling
Saul got upgraded alongside it, and this change is more meaningful than the headline suggests. In previous ChatGPT versions, instant mode and thinking mode behaved almost like two separate personalities: different tones, different formatting, and switching between them felt like swapping models entirely. Now Saul handles everything uniformly with fully adjustable speed, where the only difference is how long it reasons. The disjointed feeling is gone. The five-level reasoning slider that was previously exclusive to ChatGPT Work is now in the standard ChatGPT interface for everyone.
The most immediately obvious change is that Saul stopped rambling. OpenAI's own example: someone asking whether biking from Mission to the beach after work would leave them soaked. The old instant mode produced a wall covering rainfall, wind speed, temperature, sea fog, beach hazard warnings, plus a note that a cotton t-shirt might feel sticky. The new Saul opens with the conclusion: "You won't get soaked. The real issue is the headwind. West wind at 10 to 20 mph. Bring a thin windbreaker." When the user follows up with "I'll leave at 5:30," the old version repeated the entire forecast. The new one only updates the conclusion that changed.
Factual accuracy improvement across the GPT 5.6 family vs GPT 5.5 Instant. Grading was strict: a single factual error anywhere in a response marks the entire response incorrect. Source: OpenAI.
The bigger update is underneath: factual accuracy across the whole GPT 5.6 family improved substantially, tested on finance, healthcare, and law, the three domains where factual errors are least acceptable. The grading was strict. A single factual error anywhere in a response marks the entire response incorrect. Under that bar, Saul's error rate came in 68% lower than GPT 5.5 Instant, with Luna 62% lower.
08Astra: OpenAI's next flagship is already at release candidate
The free tier was the appetizer. Industry insider LEO broke the news that OpenAI is preparing to launch its next-generation flagship next week, codenamed Astra. It is a completely new pre-trained model, the largest OpenAI has trained since GPT 4.5. The latest internal checkpoint, codenamed mu4, has already reached release candidate status, the final version before official release.
The trail was there all along. On July 30th, an OpenAI preview video was deleted quickly, but users grabbed a screenshot showing the word "MU3" flashing on screen. This week, more leaks confirmed MU4 was in internal testing. Counting from MW to MU2 on upward, the series has quietly reached its fourth form.
The funniest part is that OpenAI announced it in the open and nobody caught it. On August 1st, they published a mathematics blog post stating their model had solved 10 open mathematical problems unsolved for more than a decade. Buried in it was a line crediting an internal build of Astra, described as their next major model.
Astra leak timeline: from MU3 screenshot to expected launch. OpenAI effectively confirmed Astra in a mathematics blog post on August 1st. Sources: leaked screenshots, OpenAI blog, industry insider LEO.
Current leaks position Astra for long-duration multi-agent collaboration, where multiple AI instances work together for hours or even days on one large complex problem. Two rumored specs are circulating. Some media reports say twice the size of GPT 5.6 Saul. Developer Haider estimates GPT 4.5 at around 5 trillion parameters and puts Astra at 7 to 10 trillion, massive either way.
With stronger infrastructure, optimizations, and a large amount of new compute coming online this year, OpenAI's service cost for running Astra may actually land below what it costs Anthropic to run Mythos 5. Haider's read on why this is the moment: OpenAI has the strongest post-training capability in the industry. That is how GPT 5.5 hit the performance it did. Its only real disadvantage against Mythos was the pre-training foundation. Astra upgrades exactly that shortcoming into the largest pre-training base in history. Top-tier fine-tuning on the biggest base anyone has built, aimed at one target: taking Fable 5 off the top spot.
Video reference: China Just Shocked Everyone With a 10 Trillion Parameter AI Model — AI Revolution. Metadata verified with YouTube oEmbed API on 2026-08-08; the displayed view count changes over time and is not used here.
References
- China Just Shocked Everyone With a 10 Trillion Parameter AI Model — AI Revolution
- Financial Times — ByteDance 10 trillion parameter model report
- Reuters — ByteDance training report coverage
- Meta — Muse Code and Muse Spark 1.2 announcement
- OpenAI — GPT 5.6 Luna free tier and Saul upgrade
- OpenAI mathematics blog post (August 1, 2026) — Astra confirmation
By N43 and Hermes for Sailor Bob News.
