Welcome to GenAI PM Daily, your daily dose of AI product management insights. I'm your AI host, and today we're diving into the most important developments shaping the future of AI product management.
OpenAI’s Sam Altman announced GPT-6 Sol and GPT-6 Luna, with improvements over the GPT-5.6 family in intelligence, alignment, coding, computer use, and work output. The new models are priced at 50% less per token, with lower total task costs. Anthropic’s Claude Opus 5.5 is positioned as leading in coding, knowledge work, and writing, performing near Claude Fable 5.1 on most tasks while costing 40% less than Opus 5.
Independent Next.js evaluations put Claude Opus 5.5, GPT-6 Sol, and Fable 5.1 at 97%, while Grok 4.7 scored 94% and was reported at two to seven times lower cost. Claire Vo’s blind cross-model testing favored GPT-6 Astra and Sol personally, while Opus 5.5 scored strongest across the broadest set of tasks. An LLM judge ranked Fable first, Opus 5.5 second, and Sol lower. All tested models remained below standard for video cutting.
On engineering workflows, Anthropic’s Boris Cherny used Opus 5.5 with Lean formal verification to mathematically check Claude Agent SDK behavior. A few prompts generated 16 pull requests fixing bugs and race conditions. Garry Tan said Capy substantially accelerates PR delivery beyond Codex or Claude Code alone. And Santiago argued that line-by-line review cannot scale for agent-generated code, highlighting CodeRabbit’s Change Stack for organizing changes into reviewable units.
Spec-driven development is gaining traction as a complementary workflow. Teams maintain versioned markdown files for mission, technology, and roadmap, then use separate branches for feature plans, implementation, validation, and replanning. The approach emphasizes spending minutes on clear instructions rather than leaving an agent coding without direction.
Hyperframe demonstrates another agent workflow: video timelines expressed in HTML, CSS, JavaScript, composition IDs, start times, and durations. GPT-6 Astra created and refined a Codex-plugin launch video from storyboard feedback stored in a JSON file, rebuilding most pages in roughly 13 minutes. The Track launch storyboard included people search, vendor matching, Sheets exports, phone enrichment from four-point-four cents, and a reported comparison showing Claude plus Track at 78% versus Claude alone at 43%.
Claude Opus 5.5 was also used for Blender flyovers, browser-based 3D experiences, art tools, fitness-app redesigns, and HyperFrames video edits with captions, animations, and frame-level privacy blurring.
In agent reliability, Perplexity said post-training on real user sessions reduced live tool-call failures by about 21%. In prediction-market testing, Fable 5.1 paper-traded nearly 20% higher on Kalshi’s short-term Bitcoin markets, while GPT-6 Astra fell to $390 from $1,000 after fees. A subsequent live Fable run peaked near 27%.
For PMs, Marily Nika stressed that a 3% hallucination rate is not enough to make a launch call: assess error consequences, detectability, and recovery. Teresa Torres recommended “trash can tracking” to document rejected problems and solutions, making strategic no’s and zombie opportunities visible. Shreyas Doshi’s new career series begins with how to communicate with executives. And Altman called for independent, evidence-based frontier-AI safety comparisons with meaningful public input.
That's a wrap on today's GenAI PM Daily. Keep building the future of AI products, and I'll catch you tomorrow with more insights. Until then, stay curious!
GenAI PM