The Product Desk

Three Frontier Models, One Week, Radically Different Economics

GPT-5.5 launched at $5/$30 per million input/output tokens — exactly double GPT-5.4's pricing. Within hours, DeepSeek shipped V4-Flash at $0.14/$0.28 under MIT license with a 1M-token context window. Z.ai's GLM-5.1 quietly topped SWE-Bench Pro at 58.4% (beating GPT-5.4 at 57.7% and Claude Opus 4.6 at 57.3%) at $1.40/M input tokens, also MIT-licensed. The result: a 35x cost spread between proprietary frontier and open-weight frontier-adjacent at increasingly comparable quality.

If you're running any high-volume AI feature on a single closed frontier model without tiered routing, you're burning 97% of your inference budget on tasks that don't require it.

The Intelligence-Per-Dollar Reframe

OpenAI's Noam Brown is pushing 2D intelligence-per-dollar charts as the new evaluation standard, and the data supports it. GPT-5.5 medium matches Claude Opus 4.7 max at 25% of the cost (~$1,200 vs ~$4,800 on Artificial Analysis benchmarks). Gemini 3.1 Pro Preview matches both at ~$900. GPT-5.5 also uses significantly fewer tokens per task than its predecessor — meaning effective cost improvement exceeds the sticker price increase. DeepSeek V4's hybrid attention architecture cuts KV cache usage to 10% of the previous generation, making that 1M context window practical at production scale.

Where Sources Agree — and Diverge

Across 17 sources, there is unanimous agreement that multi-model architecture is now table stakes. However, sources diverge sharply on DeepSeek's production viability. Technical sources validate V4-Flash for commodity workloads (summarization, classification, search ranking), noting vLLM and SGLang shipped day-0 support. But geopolitical analysts flag real risk: DeepSeek is raising at $20B+ from Tencent and Alibaba, the House Foreign Affairs Committee is advancing a distillation blacklist bill, and V4-Pro is capacity-constrained until Huawei Ascend 950 clusters ship in H2 2026. For regulated industries, 'we run on a Chinese AI model' is a procurement conversation you need to prepare for.

The Architecture You Need Now

The winning pattern is a three-tier routing architecture: (1) DeepSeek V4-Flash or GLM-5.1 for commodity inference at $0.14–$1.40/M, (2) GPT-5.5 standard at $5/$30 for general-purpose tasks, (3) Claude Opus 4.7 or GPT-5.5 Pro at $30/$180 for complex reasoning requiring maximum capability. Together AI's inference volume grew 10,000x year-over-year (30B to 300T tokens/month), confirming that AI features are moving to production scale across the industry. Your architecture must scale with demand without locking you into a single provider's pricing curve.

The GPT-5.5 API Caveat

GPT-5.5 is already live in ChatGPT and Codex, but API access is delayed pending additional safeguards. OpenAI classified GPT-5.5 as 'High' risk — meaning it could amplify existing pathways to severe harm. Do not plan hard launches around GPT-5.5 API availability until access is confirmed. Use Gemini 3.1 Pro or your existing stack as the fallback.

Four Platforms, One Week, One Message: Chatbots Are Dead

In a 48-hour window, OpenAI pivoted Codex from a coding tool into a consumer/enterprise superapp with browser control, Sheets/Slides manipulation, PDF/Docs handling, OS-wide dictation, and auto-review — then shut down Prism and folded everything into Codex. Microsoft made Copilot Agent Mode default-on for all 365 Copilot and Premium users across Word, Excel, and PowerPoint. Google launched the Gemini Enterprise Agent Platform with persistent memory and cryptographic agent identities. Anthropic put filesystem-based persistent memory into public beta for Claude Managed Agents, with scoped permissions (read-only org-wide + read-write per-user), full audit logs, and API-managed exportable/redactable memory files.

If you're a PM writing PRDs that describe AI as 'the user types a prompt and gets a response,' you're designing for an interaction model that four of the five major platforms just deprecated.

The Platform Lock-In Play

Sam Altman is explicitly framing OpenAI as an 'AI inference company,' not a model company. Greg Brockman described combining ChatGPT + Codex + AI browser into a unified enterprise superapp. This is the AWS-to-application-layer playbook: first provide infrastructure, then notice which apps are popular, then build those apps and bundle them. If your product automates CRM data entry, financial reconciliation, content management, or QA testing, you're now competing with a horizontal agent platform backed by the world's leading model. Claude now connects to 200+ apps and chains actions across them in a single chat.

Where Your Product Fits

Sources converge on three positioning options: (1) Platform player — build your own agent ecosystem (high investment, winner-take-most). (2) Best integration — become a native skill within OpenAI/Microsoft/Google agent platforms (lower risk, platform-dependent). (3) Vertical specialist — go deep where general agents can't compete (highest defensibility, smallest TAM). The worst choice is standing still.

Microsoft's default-on gambit deserves special attention. Hundreds of millions of enterprise users will encounter agentic AI without asking for it. This sets the expectation baseline. Every enterprise buyer will now compare your AI capabilities against what they get for free in Office. Your defensibility lies in domain expertise, proprietary data, and workflow-specific trust that horizontal agents cannot replicate.

The GPTs Sunset Clock Is Ticking

OpenAI announced a future GPTs-to-workspace-agents conversion tool and carefully stated GPTs 'will stay available for now' — classic platform migration messaging. If you built custom GPTs for customers, scope the architectural differences now. Workspace agents are team-centric, cross-platform, and action-oriented — fundamentally different from GPTs' single-user paradigm. Auto-converted GPTs will underperform purpose-built agents on quality and reliability.

The Guardian Agent Pattern

OpenAI's Codex auto-review feature uses a secondary agent to quality-check the primary agent's work, reducing the approval burden on users. This directly addresses 'approval fatigue' — the #1 adoption killer for agent-powered features. If you're shipping agent features, prototype this pattern: it's the only validated approach to maintaining quality while reducing human oversight overhead.

The Valuation Flip Nobody Expected

Anthropic surpassed OpenAI on secondary markets — ~$1T vs ~$880B — the first time Anthropic has been valued higher, driven by intense investor demand and Claude Code adoption. This isn't noise: it represents sophisticated capital betting that developer experience beats model benchmarks. Yet this valuation peak coincided with Anthropic's worst product week in months.

Three Bugs, One Post-Mortem, Zero Excuses

Claude Code quality degraded across Code, Agent SDK, and Cowork simultaneously — traced to three separate changes: altered reasoning effort (lowered to reduce latency), a caching bug clearing short-term context, and strict system prompts limiting response detail. Crucially, the base API was unaffected. These were product-layer bugs, not model failures. Anthropic's post-mortem admitted expanded dogfooding was needed — their team wasn't using the product enough to catch the degradation. Fixed in v2.1.116.

If Anthropic — a trillion-dollar company — took days to trace three intersecting regressions in its own product layer, what are the odds your team catches a silent quality drop from your AI provider before users do?

This is the most important operational lesson of the week. Every model provider will ship silent quality changes. You need automated regression detection at the product layer — not just API availability monitoring — running on every provider update and alerting before users notice.

The Displacement Anxiety Your Feature Design Is Ignoring

Anthropic surveyed 80,508 workers and found that those who use Claude most — the power users getting the biggest productivity gains — are 3x more likely to fear displacement than light users, with engineers leading the anxiety. Most respondents said AI gains led to 'expanded scope and more work' — the productivity dividend is captured by organizations, not individuals.

The Product Architecture Implication

If your AI features visibly replace tasks ('AI wrote this draft'), you trigger displacement anxiety in your most engaged users. If they instead amplify user judgment ('here are 5 options based on your criteria — you decide'), you preserve agency and reduce anxiety. Your power users aren't becoming evangelists — they're becoming anxious. Frame AI as expanding capability, not replacing tasks.

The Filesystem Memory Bright Spot

Amid the quality chaos, Anthropic shipped the most enterprise-ready agentic primitive this cycle: filesystem-based persistent memory for Managed Agents with scoped permissions, audit logs, and API-managed exportable/redactable files. It maps to existing enterprise mental models — files with permissions you can inspect and control. Rakuten cut first-pass errors by 97% using it. Wisedocs accelerated document verification by 30%. If you're building enterprise agent features, evaluate this in the next sprint — it solves the persistent memory blocker that's kept most agentic AI in demo mode.