Home
Apr 24, 2026
View All

GPT-5.5 Ships, DeepSeek V4 Cuts Costs, and Linear Becomes an AI Workspace

The Short Version#

Three distinct signals today: OpenAI shipped GPT-5.5 and positioned it as a premium coding and research tool at $180/M output tokens; DeepSeek V4 arrived on Vercel's AI Gateway at a fraction of that price; and Karri Saarinen casually mentioned that Linear has become his primary AI workspace — which says a lot about where embedded AI is actually sticking.

OpenAI — GPT-5.5 Ships as the New Frontier Model#

Source: https://openai.com/index/introducing-gpt-5-5 Credibility: High (first-party announcement from OpenAI)

What happened: OpenAI launched GPT-5.5 on April 23, positioning it as their "smartest model yet — faster, more capable, and built for complex tasks like coding, research, and data analysis across tools." Lenny Rachitsky, who's been running Codex in production, published a video the same day calling it the first model to pass an intelligence test that Claude Code kept failing — and saying he's happy to pay $180 per million output tokens for the Pro tier. That's not a typo. This is a premium-tier signal: OpenAI is explicitly not racing to the bottom on price.

Key details:

  • GPT-5.5 Pro pricing: ~$180/M output tokens (Lenny's framing; confirm via API pricing page)
  • Positioned for coding, research, and data analysis "across tools" — implying multi-tool agent workflows
  • OpenAI simultaneously published a Bio Bug Bounty for GPT-5.5 ($25k reward for universal jailbreaks on bio safety risks) — they're proactively red-teaming the model at launch
  • Codex Academy content published same day: automations, plugins, skills — OpenAI is building out the enablement layer to drive adoption of Codex workflows
  • System Card published alongside announcement — following their increasingly consistent practice of pairing model launches with safety documentation

Why it matters for PMs: The pricing signal is the most important thing here. At $180/M output tokens, OpenAI is explicitly segmenting their market: commodity users go to GPT-4o or cheaper alternatives; high-stakes, high-complexity workflows go to GPT-5.5 Pro. If you're building on OpenAI APIs, this is your build-vs-buy cost curve shifting again. If you're a PM at a company using AI for coding or research workflows, the question is whether the quality delta justifies 4-6x the cost of cheaper alternatives like DeepSeek V4 (launching the same day — see below). The simultaneous Codex Academy launch is also notable: OpenAI isn't just shipping a model, they're shipping adoption infrastructure.

Critical questions:

  • What specific benchmark or workflow did GPT-5.5 pass that Claude Code failed? The Lenny framing is compelling but vague — "intelligence test" needs specificity before you can use it in a stakeholder conversation
  • At $180/M, which use cases actually pencil out? Pure coding automation with long output chains could get expensive fast
  • How does this affect OpenAI's enterprise customers who've standardized on GPT-4o? Is there a migration path or are they paying more for the same workflows?
  • The Bio Bug Bounty implies real concern about misuse — does that affect enterprise compliance conversations?

Action you could take today: Pull your last 30 days of OpenAI API costs and model distribution, then model what a 10-20% migration to GPT-5.5 Pro would cost vs. the quality gains Lenny describes. If you're not running Codex workflows yet, the Academy content is now live and structured — worth 30 minutes.

Vercel + DeepSeek — V4 on AI Gateway, Both Variants#

Source: https://vercel.com/changelog/deepseek-v4-on-ai-gateway Credibility: High (first-party changelog from Vercel)

What happened: DeepSeek V4 is now available on Vercel's AI Gateway with two variants: DeepSeek V4 Pro (optimized for agentic coding tasks) and DeepSeek V4 Flash (speed-optimized). Both default to a 1M token context window. Simon Willison flagged it this morning with the framing "almost on the frontier, a fraction of the price." DeepSeek V4 landing on Vercel's gateway the same day GPT-5.5 launched at $180/M is not a coincidence — it's the market making the cost comparison explicit.

Key details:

  • Two model variants: V4 Pro (agentic coding focus) and V4 Flash (speed/latency focus)
  • 1M token context window as default across both — significant for long-context agent workflows
  • Available immediately via Vercel AI Gateway — no separate DeepSeek account needed if you're already on Vercel
  • Willison's framing: "almost on the frontier, a fraction of the price" — consistent with DeepSeek's historical pricing pattern (significantly cheaper than US frontier models)

Why it matters for PMs: This is the build-vs-buy cost curve showing up in real time. If your team is evaluating AI coding tools or agentic workflows, today you have GPT-5.5 Pro at ~$180/M and DeepSeek V4 accessible via an infrastructure you may already be paying for. The 1M context window is the real story for agentic use cases — that's enough to load an entire codebase in context. Vercel's AI Gateway is becoming a genuine model marketplace: Kimi K2.6, GPT Image 2, and now DeepSeek V4 all landed there this month. If you're routing AI traffic through Vercel, the switching cost between models is dropping toward zero.

Critical questions:

  • "Almost on the frontier" — what's the actual quality gap on coding tasks specifically? Willison's framing implies a meaningful delta still exists
  • DeepSeek is a Chinese-owned model; for enterprise customers with data residency requirements, does running it via Vercel's gateway resolve compliance concerns or just obscure them?
  • V4 Pro vs. V4 Flash — what's the latency and cost difference? Vercel's changelog doesn't specify pricing
  • Is this truly competitive with GPT-5.5 for the complex coding tasks Lenny was describing, or is the price-performance tradeoff only favorable for simpler tasks?

Action you could take today: If you're on Vercel, DeepSeek V4 is live right now in AI Gateway. Spin up a quick evaluation against your most common AI task type — coding completion, summarization, or long-context analysis — and compare output quality at the cost difference. This takes under an hour and gives you real data for the next model-selection conversation.

Karri Saarinen (Linear CEO) — Linear Is Now His Primary AI Workspace#

Source: https://x.com/karrisaarinen/status/2047035986388156611 Credibility: Medium-High (direct post from Linear's founder/CEO, describing his own workflow)

What happened: Karri Saarinen posted on April 22 that 90% of his work-related AI use has shifted to Linear. Specific workflows he called out: pulling daily reports, reviewing user frustrations, checking project launch dates, and using Linear's coding agent to make product fixes directly. This is a product leader describing how a work management tool displaced general-purpose AI tools (ChatGPT, Claude) for his actual day-to-day.

Key details:

  • 90% of AI work use now happens inside Linear — not in a standalone AI tool
  • Specific use cases: daily reports, user frustration review, launch date checks, coding agent for product fixes
  • The coding agent piece is significant: Karri is using Linear's agent to make actual code changes, not just track them
  • This is the CEO describing his personal workflow — both a genuine signal and likely a product positioning move

Why it matters for PMs: This is the "embedded AI wins the workflow" pattern showing up clearly. Karri isn't using a better AI — he's using AI that's already inside the tool where his work context lives. Linear knows his projects, his tickets, his user feedback, his launch dates. A general-purpose AI doesn't. The insight for product leaders: the AI tools that will win long-term aren't the most capable models in isolation — they're the ones embedded in systems that already hold the user's context. This should inform how you think about where to add AI to your own product. The "daily reports" and "user frustration review" use cases are particularly interesting — those are PM jobs that Linear is now automating for its own CEO.

Critical questions:

  • Is this workflow available to all Linear customers, or is Karri describing dog-fooding features that aren't shipped yet?
  • The coding agent piece — is that Linear's native agent or is he routing to Cursor/Windsurf from inside Linear?
  • "Reviewing user frustrations" inside Linear implies Linear has ingested customer feedback data — what's the integration pattern? Is this via their API or native?
  • If 90% of AI use moves inside Linear, what happens to the general-purpose AI tools he was using before? This is the retention question for ChatGPT and Claude's standalone products

Action you could take today: Look at where your highest-frequency work tasks happen today. If you're using a standalone AI tool for tasks that a product you already use could handle (Notion AI, Linear, GitHub Copilot), run a two-day experiment: force yourself to use the embedded tool for those tasks. Track whether the context advantage outweighs the capability gap.

Quick Hits#

The Thread#

The context advantage is beating the capability advantage. This week's pattern: Karri Saarinen moving 90% of his AI work into Linear, Microsoft making Agent Mode the default in Office, OpenAI building Codex Academy to drive adoption inside existing workflows. The general-purpose AI tools are losing the daily-driver battle to embedded AI — not because they're less capable, but because embedded tools already hold the user's context. The PM implication is consistent with what we saw with Wispr Flow's team features earlier this week: the new moat isn't model quality, it's being the system of record.

Sit With This#

Karri Saarinen says 90% of his AI work has moved into Linear — pulling reports, reviewing user frustrations, using the coding agent to make product fixes. Not because Linear's AI is better than Claude, but because Linear already holds his context.

For your product: Where does your user's most valuable work context live today — and is it in your product or somewhere else? If a user moved 90% of their AI use into your product tomorrow, what context would you need to have to make that worthwhile?