Anthropic filed a confidential S-1 with the SEC on June 1, 2026, at a reported $965 billion valuation, according to Fortune. OpenAI filed its own confidential S-1 exactly one week later, on June 8, per TechCrunch.
Neither document is public. But the OpenAI Anthropic IPO race already tells API buyers something concrete about where token pricing, SKU structure, and release cadence go next.
TL;DR: Both S-1s are confidential, so every revenue and margin number you've seen is third-party reporting. The pattern from a decade of usage-priced infrastructure IPOs says headline per-token prices hold or fall while effective prices rise through premium tiers and commitment floors. The Fable/Mythos split and GPT-5.5's tiering look like that playbook already running. Your hedging window is the 6 to 12 months after listing.
A direct answer to the core question: an IPO doesn't raise your per-token price on day one. It re-anchors pricing to a gross-margin target within 6 to 12 months of the first earnings call, mostly through SKU restructuring rather than list-price changes.
Key takeaways
- Anthropic filed June 1, OpenAI June 8, both confidential under the SEC's emerging growth company process. No financials are public for either.
- The Information reports Anthropic lowered its gross margin projection even as revenue forecasts rose. Gross margin will be the dominant post-listing narrative for both companies.
- Comparable IPOs (Snowflake, Cloudflare, MongoDB, Datadog) kept headline rates flat and raised effective prices through tiers and minimums.
- The practical response costs one or two engineer-weeks: a model router, per-call cost logging, and discipline about committed-use contracts.
What did OpenAI and Anthropic actually file?
Both companies submitted draft registration statements that the SEC reviews privately. The first public version of either S-1 will hit EDGAR only days before the roadshow, likely late Q3 or Q4 2026 given TechCrunch's reporting that OpenAI may list as soon as September.
The confirmed facts are thin. The New York Times confirmed Anthropic's June 1 filing the same day, and Anthropic has reportedly selected Morgan Stanley, Goldman Sachs, and JPMorgan Chase to lead its offering. The Verge independently confirmed OpenAI's June 8 filing, and CNBC reported the company says listing timing "remains up in the air."
Everything else is unconfirmed: which company prices first, final valuation, share structure, lock-ups, and anything about post-IPO product strategy. The "race" framing itself is mostly trade-press speculation. Treat any number presented as "from the S-1" with suspicion until EDGAR has it.
What do the unit economics look like before the filings go public?
The honest summary is that no reliable gross margin number exists for either company. The Information has reported OpenAI's inference gross margin was in the high single digits in 2024, improving through 2025 as the GPT-5 family was optimized for serving cost.
For Anthropic, the same outlet reports a lowered internal gross margin projection alongside rising revenue, and has framed the margin question as a long-term AI profit problem. Higher revenue with lower margin is the combination pre-IPO analyst briefings will pick at hardest.
What's company-confirmed is the spending. OpenAI announced a 4.5 GW Stargate partnership with Oracle in July 2025. Anthropic announced up to 5 GW of new compute with AWS in November 2025 and a multi-gigawatt deal with Google and Broadcom the same year.
Read those deals as a bet that future API volume will dwarf today's, and that current token prices haven't reached steady-state margin. A company locking in multi-gigawatt capacity expects to fill it.
On the revenue side, TechCrunch reported Anthropic projects $70B in revenue by 2028, citing The Information; that's an internal projection, so weight it accordingly.
One more capital signal: Bloomberg reported in late January 2026 that Nvidia paused its planned $100 billion investment in OpenAI. When a key private capital source steps back, the IPO becomes the substitute. That supports reading these listings as capital and liquidity events first, with product strategy as a second-order effect.
How do IPOs change usage-based pricing?
Public-market scrutiny re-anchors prices to a gross-margin target within 6 to 12 months of the first earnings call. The mechanism is rarely a headline rate increase. The last decade of usage-priced infrastructure IPOs shows a consistent shape:
| Company | IPO | Headline rate | Effective price | Mechanism |
|---|---|---|---|---|
| Snowflake | Sep 2020 | Held flat ~1 year | Up on large workloads within 18 months | Capacity tiers, storage repricing |
| Cloudflare | 2019 | Kept low | Up for heavy users | Tiered, usage-bundled Workers pricing |
| MongoDB | Oct 2017 | Entry tier stayed cheap | Up at the top over time | Atlas consumption tiers added post-IPO |
| Datadog | Sep 2019 | Per-host flat | Up via product surface | New SKUs (logs, RUM, security) at higher unit prices |
| Twilio | Jun 2016 | Cut repeatedly | Discounts walked back | Margin compressed, market punished the stock |
Three things repeat across these names. Headline per-unit prices stay flat or drop in the year after listing, because companies want positive product headlines in the S-1 window.
Effective prices rise through new enterprise tiers, commitment floors, and minimum spends. And release cadence holds or accelerates for the first 6 to 12 months, because the first earnings cycles need to demonstrate velocity. The Information's Tech IPO Tracker is the best running reference as the AI names join this table.
Twilio is the cautionary case in the other direction: it cut per-message prices to chase volume, compressed its own gross margin, and got punished. Both AI labs have studied that history. Expect them to defend margin through structure rather than price cuts.
Are Fable 5 and GPT-5.5 pre-IPO margin moves?
The tiering of this month's releases lines up with the playbook above almost exactly. Anthropic released Claude Fable 5 and Claude Mythos 5 in early June 2026, and TechCrunch's June 9 coverage notes that Fable 5 is the version of Mythos the public can access today.
Anthropic's own announcement positions Mythos as the higher-capability tier limited to approved organizations.
OpenAI's structure rhymes. GPT-5.5 reached general availability in GitHub Copilot on April 24, 2026, with cheaper, faster Instant-style variants priced below the flagship on the OpenAI API pricing page.
So both companies now run the same three-layer structure heading into their S-1 windows: a fast cheap tier below the flagship to protect the developer funnel, the flagship itself for adoption headlines, and a premium tier above it where margin gets recaptured. Anthropic's release cadence reinforces the velocity signal too.
Claude Opus 4.7 went GA in GitHub Models on April 16, Opus 4.8 on May 28, and Fable 5 landed days before the filing. A major model every four to six weeks is exactly what you ship when you want the S-1 window to overlap the freshest model name.
To be fair to the counterargument: the simplest read of both IPOs is employee liquidity plus capex funding, and on that read the next 12 months of API pricing look much like the last 12. That's plausible.
But the SKU structure is already in place either way, and the structure is what moves your bill.
What this means for you
The S-1 is the prospectus; the first 10-Q is where your token price actually moves. Plan for the 6-to-12-month window after listing, and do the boring work now.
Put a router behind every model call. Wrap calls in a thin adapter that takesmodel_idas a parameter. LiteLLM gives you a uniform OpenAI-format interface across 100+ providers with cost tracking and load balancing; OpenRouter is the hosted equivalent with one key for many providers. This costs one or two engineer-weeks and turns a quarter-long migration into an afternoon.
Log cost per call, then map workloads to tiers deliberately. Run the premium tier (Mythos-class, or whatever sits above GPT-5.5) on the 10 to 20% of traffic that drives 80% of your cost, and push low-complexity traffic to the fast tier. For cross-vendor price normalization, the BenchGecko llm-pricing repo maintains weekly-updated JSON for 300+ models.
Be careful with committed-use contracts. Both vendors negotiate annual commits at real discounts, but a 12-month commit in a falling-price regime is a bet that list prices won't drop more than your discount. Commit only on workloads with well-understood unit economics; leave bursty and experimental traffic on list.
When the S-1s go public, read four things. Gross margin trajectory and whether API revenue is broken out from consumer. Customer concentration (if the top 10 customers exceed 30% of revenue, pricing power is higher than the narrative suggests). Capex drawdown rates against the Stargate and AWS commitments. And stock-based compensation as a share of revenue, the quiet margin compressor in every post-IPO infrastructure name.
Don't anchor procurement decisions on listing day. Neither company wants to surprise API customers right before pricing an IPO. The risk concentrates in the first two earnings cycles after the bell rings.
Sources
- Anthropic Files to Go Public, Setting Stage for Huge I.P.O. (NYT)
- Anthropic confidentially files for IPO at a $965 billion valuation (Fortune)
- OpenAI files confidentially for IPO, following Anthropic (TechCrunch)
- OpenAI files for IPO, following Anthropic (The Verge)
- OpenAI confidentially files for IPO (CNBC)
- OpenAI barrels toward IPO that may happen in September (TechCrunch)
- Anthropic Lowers Gross Margin Projection as Revenue Skyrockets (The Information)
- Anthropic's Gross Margin Flags Long-Term AI Profit Questions (The Information)
- Anthropic projects $70B in revenue by 2028 (TechCrunch)
- Stargate advances with 4.5 GW partnership with Oracle (OpenAI)
- Anthropic and Amazon expand collaboration for up to 5 GW of compute (Anthropic)
- Anthropic expands partnership with Google and Broadcom (Anthropic)
- Claude Fable 5 and Claude Mythos 5 (Anthropic)
- Anthropic's Claude Fable 5 is a version of Mythos the public can access today (TechCrunch)
- GPT-5.5 is generally available for GitHub Copilot (GitHub Changelog)
- OpenAI API Pricing (OpenAI)
- Nvidia Pauses Plan to Invest $100 Billion in OpenAI (Bloomberg)
- The Information's Tech IPO Tracker
- LiteLLM (GitHub)
- llm-pricing: weekly-updated LLM API pricing JSON (GitHub)