Mythos-Class Models Confirmed for Public Release in the Coming Weeks, Alongside Opus 4.8 Benchmarks
Anthropic told reporters during the Opus 4.8 release cycle that its restricted Mythos-class models will move to general availability in the coming weeks. Mythos is currently locked to roughly 50 Project Glasswing partner organisations including AWS, Apple, Google, Microsoft and IBM, so a broader rollout would put autonomous vulnerability discovery in reach of any team already cleared for enterprise-tier Claude API access. The model has reportedly already identified 23,019 high-severity software vulnerabilities inside its first month of Glasswing operation, and AISI evaluated it as capable of completing a 32-step simulated corporate network attack on three runs out of ten. If you build security or coding tools on Claude, this is the single most important access change to keep watching over the next month.
Separately, the Opus 4.8 full release note confirmed pricing at $5 in and $25 out per million tokens for the standard tier, with the Fast mode at $10 in and $50 out (three times cheaper than Opus 4.7 Fast). The alignment claim that the model is four times less likely to let code flaws pass without flagging them is the change most worth verifying in your own agent loops.
Grok Build 0.1 Lands on the xAI API as a 100-Tokens-per-Second Coding Model at $1 per Million Tokens
xAI shipped Grok Build 0.1, the model that powers its Grok Build CLI, as a public beta on the xAI API on 29 May. Headline numbers are 100-plus tokens per second, a 256k-token context window, image and text inputs, and pricing at $1 in and $2 out per million tokens. That puts it well below Opus 4.8 or GPT-5.5 on price while still claiming first-class agentic harness support in Cursor, Hermes Agent, OpenClaw, Kilo Code and OpenCode.
For developers happy to bet on throughput over absolute benchmark wins, this is the most aggressive price-versus-speed coding option to drop in this window, and it is the first time xAI has made a dedicated coding model available on the API rather than reserving it to the CLI. The announcement was published on x.ai/news before mainstream coverage.
Gemini Spark Switches On for US Google AI Ultra Subscribers as Ultra Drops 60% to $99.99
Google flipped Gemini Spark live for US AI Ultra subscribers on 29 May, just over a week after announcing it at I/O 2026. Spark is a cloud-resident personal agent that keeps running when your machine is asleep: the first time Google has shipped this pattern at consumer scale rather than as a developer-only API. The pricing move sitting underneath the rollout is arguably more important than the agent itself: Ultra dropped from $249.99 to $99.99 per month, a 60 per cent cut that brings always-on agents inside reach of individual professionals rather than just enterprise pilots.
The catch worth flagging is geographic: the EU and UK are both blocked pending an AI Act review, and most analysts are pegging Q3 for opening up. If you are in a geography where Spark is live, testing it now while the agent is in early rollout is the best way to build intuitions about where autonomous agents genuinely help versus where they add friction.
OpenAI Formalises GPT-Rosalind Access for Biodefence and Public-Health Work
OpenAI used its 29 May newsroom post to put a name and a structure around the GPT-Rosalind work that has been running with selected partners for months. The Rosalind Biodefense Program now sponsors model access for vetted developers building epidemiological modelling, screening and outbreak-detection tooling, with launch support from OpenAI's own teams. In parallel, direct GPT-Rosalind access is opening to selected US government and allied public-health agencies, plus the Coalition for Epidemic Preparedness Innovations. For practitioners weighing whether to keep biothreat workflows on smaller open models, this is the first formal route to GPT-Rosalind under a clearly scoped use case.
Reuters and The Information reported that Microsoft will use Build 2026 on 2 and 3 June to launch a suite of homegrown MAI models, including a coding-specialised model aimed at strengthening GitHub Copilot. Microsoft's own pre-Build communications cite Claude Code overtaking Copilot in enterprise developer adoption as the motivation for building its own models.
Nvidia, Microsoft and Arm posted identical "A new era of PC" teasers to X over the weekend, each pointing to Jensen Huang's GTC Taipei keynote on 31 May. Multiple leaks line up the announcement with the N1X Arm laptop chip: a 20-core CPU paired with RTX 5070-class graphics and the full CUDA stack, with Dell, Lenovo, Asus and MSI preparing first devices for the second half of 2026.
Perplexity Computer Arrives Inside Word, Excel, PowerPoint, Outlook and Teams
Perplexity made its Computer agent available inside Microsoft 365 desktop apps on 29 May, alongside refreshed in-thread source attribution and team-level usage analytics for admins. The practical implication for knowledge workers is that an autonomous third-party agent now operates inside the surface where the work actually happens rather than in a separate tab: precisely the position Microsoft is trying to hold with Copilot. Teams paying for both products will need to run a real comparison on quality and cost per task, because the overlap is no longer hypothetical.
Industry Themes
Three of the in-window items broke on official company channels before mainstream press caught them: xAI's Grok Build 0.1 announcement on x.ai/news, Google's Gemini Spark flip-on for US AI Ultra subscribers, and the coordinated Nvidia, Microsoft and Arm tease for the GTC Taipei keynote. Without the direct channel scan, all three would have surfaced 24 to 36 hours later through tech press syndication.
The coding model race has clearly become the gravity well of the moment, with xAI shipping a coding API, Anthropic confirming a Mythos timeline, and Microsoft prepping a Copilot-replacing in-house model inside the same 72-hour stretch. Expect aggressive price compression on the coding tier through Q3. The other visible pattern is the consumer agent stepping out of preview, with Gemini Spark and Perplexity Computer in 365 landing on the same Friday and pitching the same idea: a cloud-resident agent that lives where you already work.