OpenAI Spud: 48 Hours In, Still Not Live. What the Delay Means.

The window opened April 14. Polymarket called 78% probability of release by April 30. Every developer with an API key has been checking the model list at api.openai.com/v1/models since then.

April 16. Nothing.

Two days of silence after a highly anticipated release window opens is not a catastrophe — the Polymarket window is April 14 to April 30, not just April 14. But it is long enough to ask the question the AI community is circling: why has Spud not shipped, what does the delay mean, and who gains from every day it does not arrive?

The Reasons Spud Has Not Shipped

OpenAI has not commented on the delay. The company does not pre-announce release dates for safety policy reasons — stating a date creates pressure to ship before safety evaluation is complete. So the silence is expected and should not be read as evasive. But the reasons for the delay fall into four categories:

Reason 1: Safety evaluation extended. The post-training safety pipeline for a model at Spud's capability level typically runs 3-6 weeks. Pretraining completed March 24. A 3-week pipeline puts the earliest release at April 14 — exactly when the window opened. A 6-week pipeline puts the latest at May 5. If red-teaming found capability thresholds requiring additional evaluation — novel forms of deception, unexpected autonomous behaviour, new biological hazard knowledge — the pipeline extends. OpenAI's Preparedness Framework requires CEO sign-off before deployment of models that exceed certain risk thresholds. This is not bureaucratic delay — it is the designed function of the safety infrastructure.

Reason 2: Infrastructure scaling. The last time OpenAI launched a major model (GPT-5, February 2026), the API experienced significant latency issues in the first 72 hours as demand exceeded provisioned capacity. A model with the expected context extension (256K-512K tokens) and improved tool call architecture requires substantially more compute per request than GPT-5. Pre-provisioning enough inference capacity to handle the global developer demand on day one — without degrading GPT-5 service for existing users — is a non-trivial operations problem. Getting this right takes time and is worth delaying the public release for.

Reason 3: Enterprise contract coordination. OpenAI's enterprise customers on custom contracts receive advance notice and coordinated rollout plans for major model releases. If there are 50+ enterprise customers needing contract amendments to access Spud at their existing pricing tiers, legal and commercial teams need time to process those. Rushing the enterprise rollout can create billing disputes, SLA complications, and support load spikes that damage the business relationships most critical to OpenAI's $25 billion 2026 revenue target.

Reason 4: Competitive timing calculation. OpenAI has a strategic interest in shipping Spud before Google I/O (May 19-20), when Google is expected to announce Gemini 3.2 or equivalent capability improvements. Shipping too early (before safety is complete) is bad. Shipping too late (after Google I/O) gives Google's announcements the "latest and greatest" narrative. The optimal window is mid-to-late April — which means April 16-25 is actually the expected peak probability window, not April 14-15.

What Competitors Have Done During the Delay

Every day Spud has not shipped, the competitive landscape has continued moving.

Gemini 3.1 Ultra (Google): Already running with a 2M token context window — 8-16x what Spud is expected to offer. Google has been publicly demonstrating multi-modal capabilities and long-context reasoning at 2M tokens since early April. The "longest context window" narrative belongs to Google until Spud ships and can be benchmarked against real-world coherence quality.

Claude 3.7 Sonnet (Anthropic): Leads on hard coding benchmarks (LiveCodeBench) and has maintained that lead since its February launch. Every day developers spend building with Claude's tool use API is a day of stickiness that Spud has to overcome. Anthropic has also been building its MCP (Model Context Protocol) ecosystem — 97M installs — which creates integration depth that pure benchmark superiority does not immediately displace.

Meta Muse Spark: Meta's Superintelligence Labs launched its first model in early April. HealthBench scores beating GPT-5.4. Closed-source positioning competing directly with OpenAI's enterprise tier. Meta's distribution (Facebook, Instagram, WhatsApp) means Muse Spark has user-facing deployment at scale that Spud will not match immediately.

The two-day delay has not fundamentally changed competitive positions — nobody has launched a clearly superior general-purpose model that displaces Spud's expected position. But it is a reminder that the AI competitive landscape does not pause while OpenAI finishes safety evaluations.

The Polymarket Math: Still On Track

Polymarket's 78% probability of release by April 30 has not moved significantly despite the two-day delay. This is the correct read — the statistical case for a late-April release is unchanged.

The base rate reasoning: OpenAI has a consistent pattern of releasing major models within 4-5 weeks of pretraining completion. Pretraining completed March 24. The 4-week mark is April 21 — next Monday. The 5-week mark is April 28. The 6-week mark is May 5. If the post-training pipeline runs 4-5 weeks (the historical median), the peak probability window is April 21-28. OpenAI has not shipped earlier than 3 weeks post-pretraining for any recent major model.

What would move Polymarket significantly downward: no release by April 25. At that point, the Google I/O timing advantage begins to erode and the probability distribution shifts toward early May.

The signal to watch: the OpenAI model list at api.openai.com/v1/models updates before the public announcement blog post. Developers who have been monitoring this endpoint will know Spud has launched before OpenAI publishes anything on its website.

What to Do If Spud Ships in the Next 72 Hours

The day-1 test protocol from the Spud window opens post remains valid. Three tests to run immediately:

Test 1 (30 minutes): Long-context coherence — Run your highest-context production prompt (80K-128K tokens) and compare quality in the first third vs. the final third of the context. If Spud has improved coherence across the full window, the quality delta should be smaller than GPT-5's.

Test 2 (45 minutes): 8-call tool chain — Build a sequential 8-tool-call workflow with dependencies (call 3 requires output of call 1, etc). Run it 10 times. Count clean completions. GPT-5 fails reliably at calls 6-8. Spud should extend the clean zone.

Test 3 (20 minutes): JSON schema stress — Run your most complex nested JSON output schema 100 times. Count violations. GPT-5 baseline: 0.2-0.5%. Spud target: under 0.1%.

If Spud passes all three at the expected improvement margins, migration from GPT-5 to Spud is justified for long-context and agentic workloads. If the improvements are marginal, the switching cost may not be worth it for most production systems.

What a Longer Delay (Past April 25) Means

If Spud has not shipped by April 25, three things change:

Narrative shift: Coverage moves from "exciting upcoming model" to "delayed model." The AI press will run "what's taking so long" stories. These create expectations that the eventual release must exceed rather than simply meet.

Google I/O pressure: Google I/O is May 19-20. If Spud ships after May 5, OpenAI may be releasing into a window where the entire AI press is focused on Google I/O preparation. The model may be technically superior to what Google announces, but the timing gives Google the "latest news" narrative at the biggest annual developer conference.

Competitor pricing moves: Every day of delay is a day for Anthropic, Google, and Meta to offer pricing or capability incentives to developers who were planning to move to Spud. OpenAI's customer success team is aware of this and will be monitoring customer signals.

Key Takeaways

• OpenAI Spud has not launched as of April 16 — the window opened April 14; Polymarket remains at 78% by April 30; the delay is expected and within the normal 3-6 week post-training safety pipeline range
• Four possible reasons for the delay: extended safety red-teaming, inference infrastructure scaling, enterprise contract coordination, competitive timing optimisation (targeting mid-to-late April before Google I/O)
• Peak probability window: April 21-28 — the 4-5 week post-pretraining mark that matches OpenAI's historical release cadence; the April 14-16 window was always early in the probability distribution
• Competitive landscape during the delay: Gemini 3.1 Ultra's 2M context window, Claude 3.7 Sonnet's coding lead, and Meta Muse Spark's HealthBench scores are all continuing to accumulate developer mindshare
• Watch api.openai.com/v1/models — Spud appears in the model list before the blog post goes live; set up a monitoring script or check manually every morning
• Day-1 test protocol stands: long-context coherence test, 8-call tool chain, complex JSON schema stress test — run these within the first 2 hours of API access

Compare current AI model capabilities and pricing while waiting with LLM API Pricing. For the full day-1 test protocol, read OpenAI Spud release window opens — what to test.