OpenAI and Anthropic File S-1s in the Same Month, Mythos-Class Models Open Up at $10/$50, Server Memory Jumps 95% in a Single Quarter

Three lines tightened at once this week. On the capital side, Anthropic (6/1) and OpenAI (6/8) each confidentially filed an S-1 with the SEC within a single week — the frontier duopoly is heading to the public market together. Last issue we judged that Anthropic's $65B Series H was most likely its final private round before listing; that call landed within seven days. A public listing means revenue-growth pressure will keep transmitting to customers through API pricing and enterprise plans. On the capability side, Anthropic on 6/9 opened up its Mythos-class capabilities — previously restricted to a small set of organizations — as Claude Fable 5 at $10/$50 per million tokens, while retiring the Claude 4 generation on 6/15; both the release cadence and the retirement cadence are accelerating. On the China side, MiniMax M3 benchmarks itself against GPT-5.5 at 5–10% of the cost. Upstream, memory was the densest source of signals this week: contract prices for server DRAM (general-purpose server memory) rose 90–95% in a single quarter, a record; capacity for HBM (the high-bandwidth memory dedicated to AI accelerators) is booked out to 2028; and NVIDIA signed a multi-year HBM co-development agreement with SK Hynix in Seoul on 6/8. Model list prices are falling while the cost of the compute substrate is rising — the tension between those two curves will determine the real prices enterprises pay over the coming quarters.

Key Events

Anthropic and OpenAI file S-1s in the same month: frontier-model pricing enters public-market scrutiny

On June 1, Anthropic confidentially submitted a draft S-1 to the SEC, confirming Morgan Stanley and Goldman Sachs as lead underwriters the next day, with market expectations pointing to a valuation around the trillion-dollar mark. On June 8, OpenAI confirmed it had also confidentially filed an S-1, noting that timing remains undecided and may take a while. Last issue we covered Anthropic's $65B Series H and judged it was most likely the company's final private round before going public; that judgment was confirmed within seven days. With both companies reaching the filing stage in the same month, the capital structure of the frontier-model industry has formally shifted from private funding to public-market discipline.

For enterprise customers, the transmission path is concrete: a public company answers to shareholders on revenue growth every quarter, and API pricing, enterprise plan structure, and billing models are the most direct revenue levers a model vendor holds. This year's moves already point the way — GPT-5.5 doubled prices to $5/$30 per million input/output tokens, and Anthropic shifted enterprise customers from flat plans to usage-based billing. After these filings, the frequency and magnitude of such adjustments will be driven by ever-stronger revenue targets. The stability of any price terms signed today needs to be re-evaluated against the vendor's IPO cycle. AgentsFlare's multi-model routing keeps enterprises able to switch among Claude, GPT, Gemini, and open-source models at task-level granularity, so the pricing-power contest never binds you to a single vendor's capital cycle.

Sources: Anthropic files to go public — 2026-06-06; Confidential submission of draft S-1 | OpenAI — 2026-06-08; OpenAI files for US IPO | Al Jazeera — 2026-06-08

Claude Fable 5 goes GA: Mythos-class capability moves downmarket while the Claude 4 generation exits

On June 9, Anthropic released Claude Fable 5 — the first generally available model of its fifth generation, and the first time Mythos-class capability has left the restricted-access framework of Project Glasswing and opened to all customers. Mythos-class capability had previously been available only to a small set of organizations through a restricted preview, during which it demonstrated the ability to autonomously discover and chain zero-day exploits across operating systems and browsers. TechCrunch noted that the broad release came just days after Anthropic publicly warned about the risks of highly capable models — the safety narrative and the commercialization are advancing in step. On pricing, Fable 5 comes in at $10/$50 per million input/output tokens: above the standard tier of Opus 4.8 ($5/$25), but less than half of Mythos Preview ($25/$125). Pro, Max, Team, and seat-based Enterprise plans can use it at no extra cost from 6/9 to 6/22, after which it draws on usage credits. AgentsFlare has onboarded Fable 5; it is available for evaluation now.

The other half of the week's news is the exit: Claude Sonnet 4 and Claude Opus 4 (the May 2025 snapshots) retire from the API on June 15 at 9 AM Pacific, after which requests pinned to those version IDs return errors. Taken together, Anthropic's generation management runs on a cadence of pushing top capability down faster while clearing trailing versions out faster — the migration window available to enterprises keeps shrinking, with the gap from release to predecessor retirement now measured in weeks. Model lifecycle management is turning from an annual planning topic into a quarterly execution task.

Sources: Claude Fable 5 and Claude Mythos 5 | Anthropic — 2026-06-09; Anthropic releases Claude Fable | TechCrunch— 2026-06-09; Claude Sonnet 4 and Opus 4 Deprecation | MindStudio — 2026-06

GitHub Copilot moves to usage-based billing: flat subscriptions give way under frontier-model costs

As of June 1, GitHub Copilot switched from per-request to usage-based billing. Nominal plan prices stay put (Pro $10/month, Pro+ $39/month, Business $19/user/month), but every interaction is now metered at the underlying model's API rate and deducted as GitHub AI Credits; annual subscribers who stay on per-request billing face sharply higher multipliers for premium models, with Opus 4.7 rising from 7.5x to 27x. For teams that lean on Copilot Chat, agentic coding sessions, and large context windows, the cost curve has materially changed — the flat subscription used to absorb the true inference cost of frontier models, and that cost now lands directly on the end bill.

This is the most consequential billing-structure signal of the week, because it reveals a pattern that will replicate across other products: in an environment of high frontier API costs and rising upstream compute prices, flat subscription pricing is unsustainable for vendors, and usage-based billing will spread upward from the model API layer into developer tools, agent platforms, and SaaS applications. Enterprise AI spend is shifting from a predictable subscription line into a volatile usage line. June is the first full billing cycle under Copilot's new scheme, and the window to audit token consumption team by team is right now. AgentsFlare's cost attribution and quota management can decompose this volatility by team, project, and model, giving budget resets a factual basis; and once usage accumulates to scale, billing tiers tend to loosen accordingly — the price space an enterprise can actually reach is larger than the published rate card.

Sources: GitHub Copilot Ends Flat Pricing June 1 | Enterprise DNA — 2026-06; GitHub Copilot pricing change | Apidog— 2026-06

MiniMax M3: claims 5–10% of GPT-5.5's cost, but the open-source promise shrinks to open weight

On June 1, Shanghai-based MiniMax released M3, with three technical headlines: a proprietary MSA sparse-attention architecture, context up to one million tokens, and native multimodal input (text/image/video), positioned for long-horizon coding-agent workloads. The official benchmark claim is parity with or wins over GPT-5.5 and Gemini 3.1 Pro on key coding and reasoning benchmarks, at 5–10% of their cost. The release extends the cadence of China's open-source camp this year — DeepSeek V4 and Kimi K2.6 in April, MiniMax M3 in June — keeping up the densifying supply of frontier-grade open capability.

The openness level needs flagging: neither the weights nor the technical report was released at launch. MiniMax promised both within ten days on Hugging Face and GitHub, continuing its modified-MIT license, with training code and inference operators excluded. Until the weights land, none of the benchmark claims can be independently verified; by its current scope of openness, M3 qualifies as open weight, still two layers — training code and operators — short of fully open source. For enterprises evaluating self-hosted routes, the sensible move is to wait for the weights and third-party reproductions before making architecture decisions. If M3's cost claims hold up, it will join DeepSeek V3.2 in pushing down the open-source price of intelligence, adding a substantive anchor against closed-source vendors' pricing strategies.

Sources: MiniMax M3 | MiniMax Research — 2026-06-01; MiniMax-M3 debuts | VentureBeat — 2026-06-01; MiniMax M3: Frontier Claims, Unverified Benchmarks | TechTimes — 2026-06-01

OpenAI plugs into Oracle cloud commitments: cloud budgets formally become a model distribution channel

On June 10, OpenAI and Oracle announced that OCI customers will, in the coming weeks, be able to apply Oracle Universal Credits directly toward OpenAI models and Codex, with no separate procurement process. The accompanying data point: Codex now serves over 2 million weekly users, up 5x in three months, growing more than 70% month over month. The mechanism deserves unpacking: budget an enterprise has already committed to a cloud vendor can now convert directly into model and coding-agent consumption — procurement friction drops sharply, and model purchasing gets folded into the cloud contract's settlement and discount system.

This is the ecosystem-integration signal landing at the procurement layer. The model vendor gains a negotiation-free channel into enterprises; the cloud vendor binds model consumption into its own commitment pool; both sides get what they want — and the enterprise's flexibility to switch models within the contract term is the most easily overlooked cost in this structure. Before buying models with cloud credits, two things are worth confirming: first, whether the model mix can be adjusted within the contract term; second, whether model consumption can be split by team and project. The stakes of the second: credit deductions show up in the cloud bill as a single aggregate by default, with model spend blended into compute, storage, and everything else — when credits burn faster than planned, you can't locate which team or which business line consumed them, internal chargeback and budget accountability have no basis, and at renewal you can't produce your own real usage structure, leaving you to accept whatever framing the cloud vendor offers. AgentsFlare's value as an independent access layer is direct here — model choice stays in your own routing rules and never follows any one cloud's commitment structure; meanwhile, the pricing and service assurances that committed volume is supposed to buy are available on AgentsFlare at your actual usage, with the scale commitment carried for you, no need to bind yourself into any single vendor's contract first.

Sources: Access OpenAI models and Codex through your Oracle cloud commitment | OpenAI — 2026-06-10

Upstream Compute Signals

New this issue: industry-chain signals already argued by the market or by institutions, presented as signal — who's arguing it — transmission path, without rebuilding the full argument. One admission criterion: the signal must transmit to model pricing, compute availability, or enterprise deployment cost.

HBM is being locked up, one deal at a time. HBM (High Bandwidth Memory) is the dedicated memory packaged directly alongside AI accelerator chips — however strong the GPU, it idles if data can't be fed in fast enough, so a card's real inference throughput is largely set by its HBM, making it the most fought-over component in the AI compute chain. On June 8, Jensen Huang visited Seoul, where NVIDIA signed a multi-year HBM co-development agreement with SK Hynix (two years and extensible) and struck AI cloud and data center deals with SK Telecom, Naver, and Doosan Group. The backdrop: HBM capacity at the three memory makers (SK Hynix, Samsung, Micron) is essentially sold out through 2026, and multiple research houses estimate demand growing 80–100% a year against supply growth of only 50–60%, with the gap running to 2028 or beyond. Last issue we recorded the other half of the picture: Anthropic's Series H brought in Samsung, SK Hynix, and Micron as strategic infrastructure partners. Chip vendors and model vendors alike are locking up memory supply directly. HBM sets the cost floor of inference throughput, and who holds the locked-in relationships determines who gets compute, and at what price — a chain that ends at the per-million-token price of every API.

DRAM price increases have reached server and cloud bills. DRAM is the general-purpose memory in every server; it comes from the same three manufacturers as HBM and competes for the same production capacity. TrendForce data shows server DRAM contract prices rose 90–95% QoQ in Q1 (a record), with another 58–63% expected in Q2, as the big three tilt capacity toward HBM and North American cloud providers accelerate AI inference deployments while locking in supply through long-term agreements. The industry expects overall server prices to rise 5–10% between April and September, with memory-heavy configurations up more; cloud providers will absorb part of it through margin and pass the rest through via instance prices, shrinking free tiers, or surcharges. The transmission path is complete and already operating: memory → servers → cloud instances → inference cost → API and self-hosted TCO. Nominal model-API price cuts and real compute-substrate price increases are happening at the same time.

Advanced packaging remains the physical ceiling on shipment volume. An AI accelerator only ships once its GPU die and HBM stacks are assembled onto a single substrate; that assembly step is advanced packaging, today done mostly on TSMC's CoWoS lines — if chips can be made but can't be assembled, they still can't ship. CoWoS lead times run 50+ weeks by various estimates, stretching to 78–104 weeks at the high end, while capacity expands from roughly 35K wafers/month at the end of 2024 to a projected 125–130K by the end of 2026 — most of it already reserved by NVIDIA. The pace of packaging expansion sets the increment of global AI accelerator supply in 2027, and determines how much capacity second suppliers like AMD can obtain — the second compute curve we covered in last issue's AMD earnings has its physical ceiling here.

The market is already trading the bottleneck logic — signals and counter-signals coexist. Capital markets have priced these constraints heavily: the Philadelphia Semiconductor Index hit an all-time high on April 24, semiconductor ETFs were up more than 50% year-to-date as of end-May, and the argument circulating in the market runs: find the most supply-constrained link in the chain and bet on that link's suppliers holding pricing power — memory and packaging are exactly the positions named repeatedly under this frame. Counter-signals exist alongside: parts of the GPU rental market are starting to soften, and some analysts warn the sector is priced for perfection. On the regulatory side, the U.S. BIS issued new guidance on May 31 extending advanced-AI-chip export-license requirements to entities outside China whose headquarters or ultimate parent sits in China (including Macau), closing the third-country procurement route; the corresponding move on the China side is domestic compute share continuing to climb, with IDC and others projecting Huawei at roughly 50% of China's AI chip market in 2026. Our reason for recording these signals is that their transmission paths to the model layer are real; judgments at the stock-price level sit outside this column's scope.

Sources: NVIDIA South Korea AI deals — 2026-06-08; AI Server Demand to Drive Memory Contract Price Increases in 2Q26 | TrendForce — 2026; DRAM prices predicted to jump 63% in Q2 | Tom's Hardware — 2026; TSMC Foundry Allocation Status Q1 2026 | Silicon Analysts — 2026; US upgrades AI chip export controls | Sina Tech — 2026-06-01

This Week's Structural Read

Put the model layer and the upstream on the same chart and this week shows two cost curves moving in opposite directions. The unit price of intelligence at the model layer is falling: Mythos-class capability dropped from $25/$125 to $10/$50 and opened to everyone, and MiniMax M3 claims flagship-adjacent performance at 5–10% of the cost. The cost of the compute substrate is rising: DRAM up 90–95% in a quarter, the HBM gap stretching to 2028, server prices expected up 5–10% in the second half. The two curves won't simply cancel out; the tension will release through more frequent changes to price structures — cuts, hikes, and billing-model switches arriving within the same quarter. Copilot's move from flat subscription to usage-based billing is the first clear sample, and a post-filing OpenAI and Anthropic have stronger motives to follow with adjustments of their own. What enterprises face in H2 2026 is a price sheet with rising volatility, and budget management needs to move from annual lock-in to quarterly rolling review.

The other supply-chain shift is lock-in relationships extending upstream. Model vendors are taking direct stakes in or binding themselves to memory makers (Samsung/SK Hynix/Micron in Anthropic's Series H); chip vendors are signing multi-year agreements with memory makers (NVIDIA and SK Hynix); cloud vendors are binding model distribution into their commitment pools (Oracle and OpenAI). Every layer is trading long-term contracts for certainty, and the cost of those contracts will eventually settle into every price an enterprise procures at. Who has locked up whom along the supply chain determines the path and speed of price transmission for the next two years.

What to Do This Week

Five actions have explicit time windows tied to this week's events. First, before June 15, inventory every service calling Claude Sonnet 4 / Opus 4 by pinned version ID — after retirement those requests fail outright, which is production-incident-grade risk. Second, finish evaluating Fable 5 before the free window closes on June 22, but run budget math at $10/$50; any cost commitment based on free-window experience will mislead. Third, June is the first full billing cycle under Copilot's usage-based scheme — at month end, audit token consumption structure team by team and reset quotas, with particular attention to the share consumed by long agentic sessions. Fourth, price the memory and server increases into your H2 2026 self-hosted inference TCO and recompute the break-even point between API and self-hosted — upstream price rises erode the self-hosted route more directly than the API route. Fifth, before procuring OpenAI through Oracle credits, confirm the contract's terms for adjusting the model mix and the attribution granularity of consumption data.

Models are getting faster, prices are in motion, and the upstream is being locked up; the one thing an enterprise fully controls is its own access layer. Keep routing rules, cost attribution, and switching capability in your own hands, and price volatility stays at the level of parameter tuning, never escalating into architecture rework.

References

Anthropic / OpenAI S-1

Confidential submission of draft S-1 to the SEC | OpenAI — 2026-06-08

Tech giant OpenAI files for US initial public offering | Al Jazeera — 2026-06-08

Anthropic files to go public in a potentially trillion-dollar debut | Pearson Finance News — 2026-06-06

Claude Fable 5 / Claude 4 retirement

Claude Fable 5 and Claude Mythos 5 | Anthropic — 2026-06-09

Anthropic releases Claude Fable, a version of Mythos | TechCrunch — 2026-06-09

Anthropic brings Mythos to the masses with Claude Fable 5 | VentureBeat — 2026-06-09

Claude Sonnet 4 and Opus 4 Deprecation | MindStudio — 2026-06

GitHub Copilot billing switch

GitHub Copilot Ends Flat Pricing June 1 | Enterprise DNA — 2026-06

GitHub Copilot pricing change | Apidog — 2026-06

MiniMax M3

MiniMax M3: Frontier Coding, 1M Context, Native Multimodality | MiniMax Research — 2026-06-01

MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance | VentureBeat — 2026-06-01

MiniMax M3 Open-Weight Coding Model: Frontier Claims, Unverified Benchmarks | TechTimes — 2026-06-01

OpenAI × Oracle

Access OpenAI models and Codex through your Oracle cloud commitment | OpenAI — 2026-06-10

Upstream compute signals

NVIDIA Secures Six Deals with AI Tech Giants in South Korea | AI Magazine — 2026-06-08

Nvidia locks down Korean memory deals as HBM shortage runs through 2028 | gagadget — 2026-06

Samsung and SK hynix warn AI-driven memory shortages could last until 2027 and beyond | Tom's Hardware — 2026

AI Server Demand to Drive Memory Contract Price Increases in 2Q26 | TrendForce — 2026-03-31

DRAM prices predicted to jump 63% in Q2, NAND up to 75% | Tom's Hardware — 2026

Server Price Increases in 2026 | Hostkey — 2026

TSMC Foundry Allocation Status Q1 2026: N3 Fully Booked, CoWoS 50+ Weeks | Silicon Analysts — 2026

The Great Packaging Pivot: How TSMC is Doubling CoWoS Capacity | FinancialContent — 2026

Semiconductors in 2026: The AI-Driven Upswing Meets Structural Bottlenecks — 2026

Data-Center Power Constraints: The Critical AI Growth Threat | SquaredTech — 2026

US upgrades AI chip export controls, closing the offshore-subsidiary route | Sina Tech — 2026-06-01

US Commerce Department issues new rule on AI chips to Chinese firms outside China | VOA — 2026-06-01