GPT-5.6 Luna is now the default on ChatGPT Free and Go with text-chat caps removed, while Sol moves up to Plus and Pro claiming 40% fewer reasoning errors. The same day, OpenAI called in the APA.
OpenAI made GPT-5.6 Luna the default for Free and Go accounts and removed text-chat limits, pushing an improved Sol to Plus and Pro. OpenAI says responses containing at least one factual error were about 62% less common with Luna and 68% less common with Sol than with GPT-5.5 Instant, in its August 6 announcement.
OpenAI says responses with at least one factual error were about 68% less common with the updated Sol than with GPT-5.5 Instant, and the GPT-5.6 family carries a 1.05M-token context window. Meta answered with Muse Spark 1.2 on August 5, closed weights behind its developer API, less than a month after 1.1 on July 9.
The Trump administration imposed a 15% tariff on polysilicon derivative products this week, reportedly paired with minimum import prices. Wafers and solar cells are the direct hit, and datacenter power budgets sit downstream of both.
A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.
Join the founding list →OpenAI made GPT-5.6 Luna the default model for ChatGPT Free and Go accounts and removed text-chat limits on those tiers, while pushing an improved GPT-5.6 Sol to Plus and Pro users. OpenAI says responses with at least one factual error were about 68% less common with the updated Sol than with GPT-5.5 Instant, and the GPT-5.6...
Continue reading arrow_outwardGPT-5.6 Luna is now the default for ChatGPT Free and Go with text-chat limits removed, while an improved Sol goes to Plus and Pro claiming 68% fewer factual errors than GPT-5.5 Instant and a new reasoning slider.
Iran and Oman advanced a draft Hormuz transit arrangement that would separate inbound and outbound lanes and bar US and Israeli-linked vessels. Netanyahu warned on August 5 that Israel could unilaterally resume strikes on Iran's nuclear program.
Mood is Fear: SpaceX's debut earnings print sank on capex guidance despite Musk's $1 trillion revenue target, and Washington's new 15% polysilicon derivatives tariff raises input costs across chip substrates and solar.
OpenAI removes usage caps for the tier least able to pay and simultaneously partners with the American Psychological Association: the safety announcement and the growth lever landed together, with concrete commitments still unspecified.
Making Luna the default on Free and Go, with no text-chat limits, converts OpenAI's cheapest tier into its widest distribution channel. Sol's gains, about 68% fewer factual errors than GPT-5.5 Instant, stay behind the Plus and Pro paywall, so the free tier competes on volume while the paid tier competes on capability.
Meta shipped Muse Spark 1.2 on August 5, less than a month after 1.1 on July 9, and distributed it through its developer model API rather than as open weights. Benchmark disclosure remains thin, underlining the split between Llama's open releases and Meta's closed generation stack.
The 15% tariff on polysilicon derivative products, reportedly paired with minimum import prices, targets an upstream input for wafers and solar cells. The stated goal is supply-chain security, but the cost travels downstream into chip substrates at the moment capital markets are already punishing capex guidance.
The 80% Cut Heard Everywhere
Read full analysis arrow_outward
Deep Dive
7.2/ 10
Google shipped an editor that assumes you are managing agents, not typing code, and it is free while the credit rules are still being written.
Antigravity is for developers who already trust agents with multi-step work and want a command center for running several of them in parallel instead of babysitting one chat. Worth adopting now as a second editor on a staging-scoped project, not as the place your production deploy keys live. If your workflow is single-session, one-file-at-a-time coding, Cursor or Windsurf will feel less like overhead.
“Reviewers keep coming back to the manager view and the agent-first workflow as the real differentiator, especially when a task spans code, tests, and browser checks.”Nimbalyst review
Copilot Workspace is generally available and is built for autonomous multi-file edits and repository-wide migrations rather than single-function completions. The new Autopilot and Fleet modes are the reason to revisit it: point them at a framework upgrade or a mechanical refactor that spans dozens of files, then review the diff as a unit. Treat the first run as a scoping exercise on a branch you are willing to throw away.
Try it arrow_outwardThe Iris.ai Researcher Workspace is a software suite for analyzing and summarizing bodies of research data, aimed at literature reviews and systematic screening rather than one-off questions. It earns its place when you have a corpus of dozens or hundreds of documents to triage and need consistent extraction across all of them. Use it to narrow the reading list, then read the ten papers that survive yourself.
Try it arrow_outwardAI Reality builds simple web-based augmented reality prototypes using Stable Diffusion and OpenAI models to generate the assets, so you can put a rough AR concept in front of a stakeholder without a Unity project. This is prototype fidelity, not shipping fidelity: use it to test whether an AR idea is worth budgeting before anyone commits engineering time. Expect to rebuild anything that survives the pitch.
Try it arrow_outwardMarTech's running AI release roundup, updated August 6, is a dated log of what marketing and CX vendors actually shipped, including Klaviyo's agreement to acquire customer success startup Agency, which raised $32 million from Sequoia, Menlo Ventures and Felicis, with founder Elias Torres joining as chief product officer and a 25-person team folding into Klaviyo's Composer and Customer Agent products. Skim the dated entries rather than the vendor blog posts: acquisitions and staffing moves tell you which features will survive the next roadmap cut. Bookmark it as a weekly ten-minute read.
Try it arrow_outwardRe-benchmark your prompt stack against Luna. If your product routes free users to a model, the default just changed. Run your ten hardest evals against Luna and Sol side by side and record where the 68% factual-error claim actually shows up in your domain.
Put Copilot Workspace on a real backlog item. It hit general availability with Autopilot and Fleet modes. Give Fleet three independent bug tickets in parallel and measure how many land as clean PRs without human edits.
Test Antigravity on a deploy, not a demo. It is a VS Code fork where the agent plans, tests, and deploys. Point it at a low-stakes service, watch the plan it writes before execution, and log every step you had to override.
OpenAI just made its cheapest tier its widest channel, and unlimited text chat on Free is a harder moat than any eval table. Meta went the other direction with Muse Spark 1.2, closed weights and thin published numbers, which tells you it is competing on product surface rather than developer goodwill. Meanwhile the APA partnership lands the same day the caps come off, which is what shipping under scrutiny now looks like. Build for the model your users get for free, because that is the one that will define their expectations of yours.
AI analysis, world news, markets, and tools. One briefing, delivered free.