OpenAI's GPT-Image-2 launch dominated the past 48 hours, claiming the #1 spot on the Arena text-to-image leaderboard by a wide margin and setting a new benchmark for photorealistic generation with multi-pass self-review. But the more consequential story sits underneath the image hype: Anthropic's unreleased Mythos model was accessed by unauthorized users (Bloomberg), Mozilla used a Mythos preview to find 271 vulnerabilities in Firefox 150, and Anthropic botched a Claude Code pricing change—briefly moving it from the $20 Pro tier to $100+ Max before reverting. Three separate Anthropic storylines in 48 hours, none of them planned announcements.
Bloomberg reports that Anthropic's unreleased Mythos model has been accessed by unauthorized users, raising immediate questions about pre-release access controls at one of the industry's most safety-focused labs. Separately, Mozilla confirmed using a Mythos preview to identify 271 vulnerabilities in Firefox 150, confirming the model exists and is already in enterprise security production. The gap between "unauthorized access" and "actively deployed in browser security" suggests Mythos is further along than Anthropic has publicly acknowledged.
Why it matters
A safety-first lab losing control of an unreleased frontier model is an embarrassing contradiction—and the Firefox security use case confirms Mythos is a serious capability jump, not a minor iteration.
OpenAI launched GPT-Image-2 with a self-review iteration mechanism that runs roughly 11 minutes of multi-pass refinement, producing outputs users and benchmarks are calling the new photorealism standard. The model tops the Arena text-to-image leaderboard with a significant gap over the next competitor, and independent testers confirm strong performance on cinematic color coherence, complex compositions, and documentary-style fidelity. Simon Willison's raccoon-with-ham-radio benchmark also showed clear improvement over its predecessor.
Why it matters
Image generation has been a crowded, commoditized space—a convincing leaderboard lead by OpenAI resets the competitive baseline and puts pressure on Midjourney, Stability, and Google's Imagen stack simultaneously.
Across 880 evaluations, Claude Haiku 4.5 equipped with a single agent skill scored 84.3% versus baseline Opus 4.7's 80.5%—a smaller, cheaper model beating Anthropic's flagship through scaffolding alone. The test covered nine models and suggests that agentic capability layering is now a more efficient lever than raw model scale for many production workloads.
Why it matters
This directly challenges the "buy the biggest model" default for enterprise deployments and strengthens the case that agent framework design, not model selection, is the primary cost-performance variable going forward.
Anthropic quietly removed Claude Code from the $20/month Pro plan feature list, with support docs updated to reference Max plan only—implying a jump to $100+/month. After user backlash and community confusion, Anthropic appeared to partially reverse course. GitHub Copilot simultaneously tightened its Individual plan limits and restricted Claude Opus 4.7 access to its $39/month Pro+ tier, citing agentic workflow compute demands.
Why it matters
Two platforms in 48 hours repricing agentic compute upward signals that the economics of running autonomous coding agents at scale don't work at consumer price points—expect more tier restructuring across the industry.
Mozilla deployed Claude Mythos Preview in a production security pipeline and identified 271 vulnerabilities in Firefox 150, according to a quote from Bobby Holley flagged by Simon Willison. This is a concrete, named production deployment of a pre-release frontier model doing high-stakes security work on one of the world's most widely deployed browsers.
Why it matters
271 vulnerabilities in a mature, heavily audited codebase is a striking number—it validates AI-assisted security auditing as a genuine capability step, not a demo, and sets a benchmark other labs will need to match.
Claude Code's Auto mode allows tool execution without per-call user approval, and the dangerouslyDisableSandbox flag can be invoked without explicit operator configuration, enabling arbitrary Bash command execution in production environments. Users are documenting this as an undocumented risk surface in agentic deployments.
Why it matters
As Claude Code gains enterprise adoption, unsandboxed autonomous shell access is a material security risk—this will force Anthropic to clarify guardrails and likely prompts enterprise buyers to audit their current Claude Code deployments immediately.
Agentic compute costs are being repriced upward simultaneously
Anthropic (Claude Code to Max tier), GitHub (Copilot Pro+ gating Opus 4.7), and Google (Ads Advisor agentic features behind enterprise tiers) all moved in the same direction within 48 hours.
Anthropic's unreleased models are already in active enterprise deployments
Both the unauthorized Mythos access story and Mozilla's Firefox security use case confirm pre-release frontier models are operating in production well before public announcements.
Local and open-weight model momentum continues
IBM Granite 4.1-8B dropped on HuggingFace, Unsloth shipped Kimi K2.6 GGUF quantizations, and the Kimi K2.5→K2.6 MineBench comparison shows meaningful capability gains at sub-$3 inference costs.
Signal vs noise
Signal · underreported
The Haiku 4.5 + agent skill outperforming Opus 4.7 finding is underreported and structurally important. If agent scaffolding reliably beats raw model scale at 5x lower cost, it reshapes enterprise AI procurement logic and undercuts the premium tier pricing story that Anthropic, OpenAI, and others are simultaneously pushing.
Noise · overhyped
The GPT-Image-2 photorealism coverage. Real-world image generation quality is now good enough across multiple providers that marginal leaderboard gaps don't translate to meaningful production differentiation for most use cases—and the 11-minute generation time for multi-pass refinement limits practical deployment scenarios considerably.
What to watch
Track whether Anthropic formalizes Claude Code pricing and sandboxing policy next week—both the tier restructuring and the dangerouslyDisableSandbox exposure need official responses, and how they handle them will signal how seriously they're treating enterprise security at scale.