Anthropic's Opus 5.5 Hits Fable 5.1's Bar at 40% of the Price — Then OpenAI Cuts GPT-6 in Half the Same Day

Anthropic shipped Claude Opus 5.5 on Sept 22, 2026, priced at $4 per million input tokens and $20 per million output tokens — a 40% cut from Opus 5 ($5/$25), which had only launched60 days earlier on July 24. The release note is unambiguous on the positioning: Opus 5.5 matches Claude Fable 5.1's frontier performance at a meaningfully lower price point, and on most agentic benchmarks the cheaper model beats its premium stablemate.

OpenAI didn't wait. GPT-6 Sol and GPT-6 Luna hit the API on Sept 22, 2026, around 18:00 UTC — the same afternoon as Opus 5.5. Pricing landed at $2 input / $10 output per million tokens for Sol and $0.10 / $0.50 for Luna. That cuts Sol's promotional GPT-5.6 Sol rate ($4/$20) by another 50% on input and output, and brings the cache-read rate to $0.20/M. A customer moving a frontier workload from Fable 5.1 to GPT-6 Sol today pays roughly one-fifth of what they paid in June.

The headline efficiency claim — Claude Fable 5-level scores at roughly one-fifth the cost — makes Fable 5.1's $10/$50 price point structurally indefensible at the enterprise level. OpenAI is also explicitly leaning into the breadth-of-portfolio argument: GPT-6 Astra at the top, Sol for production, Luna for volume — each rung cheaper than Anthropic's closest equivalent.

Reading the chart in chronological order, the headline price for a frontier-class query fell from $60 per million mixed tokens at the Fable 5 launch on June 9 to $12 for GPT-6 Sol on Sept 22 — a five-fold compression in fifteen weeks. Anthropic's cheaper Opus 5.5 resets the second tier to $24; OpenAI's Luna tier takes the same workload down to $0.60. This is not gradual deflation. This is a controlled demolition of the price floor that both labs' IPO pitches were originally written against.

The price war lands directly on the IPO calendars. Anthropic told investors in mid-August2026 that its annualized run-rate revenue hit $65 billion at the end of July, with the company projecting $100–120B by year-end. Six Anthropic investors told the Financial Times they expect an October 2026 listing at $2 trillion or more — up from the $965 billion post-money set in its May 2026 Series H. At30× projected forward revenue, the conversation has already drifted to $3 trillion.

OpenAI confidentially filed its S-1 on May 22, 2026, made its public S-1 public in August, and as of early September was sitting on $40B annualized revenue — still ahead in scale but slipping behind on growth. CFO Sarah Friar said Aug 19, 2026 that OpenAI will list in 2027, or sooner if it continues to inflect. The pre-IPO valuation target remains $1 trillion or above, against a last-private mark of $852B from the March 2026 round. Sam Altman has confirmed the IPO will not happen in 2026, citing both AI-safety optics and the changed competitive backdrop.

Both labs now have to defend those numbers against the pricing curve that was reset on Sept 22. OpenAI's $1T target on $40B ARR is roughly a 25× revenue multiple — defensible only if revenue keeps compounding at the 140% YoY pace of 2025–2026. Anthropic's $2T target on a $65B run rate is closer to 31×, defensible only if its $100–120B 2026 ARR target lands. A second wave of token-price compression inside Q4 — which the Sept 22 events make likely — would force both sets of underwriters to rebuild their pitch.

The price collapse isn't a margin problem for the silicon vendors — it's a volume catalyst. Two earnings prints bracketed the Sept 22 release and confirm the demand side is not slowing down. NVIDIA's Q2 FY2027 (quarter ended Jul 26, 2026) posted $96.2B in revenue, up 106% YoY, with Data Center revenue of $89.0B, up 117% YoY, on Blackwell Ultra ramp. Broadcom's Q3 FY2026 posted $29.6B in revenue and management guided AI-semiconductor revenue to $115B in FY2027 — roughly doubling from FY2026 — and $230B in FY2028, anchored by the 3.5 GW of Google TPU capacity Broadcom is co-building for Anthropic starting in 2027.

The supply chain runs through four choke points, and each one has a public beneficiary. NVIDIA supplies the GPUs that train and run both labs — Anthropic's $80B-plus cloud-spend projection through 2029 includes BigCompute on Azure ramps and NVIDIA silicon in the non-Google portion of its fleet. Broadcom co-designs the Google TPU silicon that anchors Anthropic's 3.5 GW from 2027, with $128B of purchase commitments already disclosed and $55B of deliveries scheduled for FY2027 alone. TSMC fabricates both the NVIDIA Blackwell and the Broadcom-Google TPU at 3nm and below. CoreWeave sits on a $22.4B OpenAI contract plus an Anthropic deal from April 2026, taking a chunk of AI cloud spend that doesn't flow through the hyperscalers.

The non-obvious read: the labs are trading margin for share in a market whose dollar volume is still accelerating, and the chip complex — not the labs — is the structural winner of the AI capex cycle. Anthropic's IPO at $2T and OpenAI's at $1T+ both assume the labs hold the pricing power; the Sept 22 reset suggests they don't. The silicon vendors, in contrast, are selling into an inelastic demand curve where every 10% cut in token price roughly doubles the volume of queries that the same dollar of compute can absorb.

If Anthropic lists at $2T in October and OpenAI at $1T+ in 2027, the cloud partners each lab anchors on have a very different revenue path. Amazon's relationship with Anthropic is the most monetizable: $25B in cumulative Amazon equity into Anthropic (per the April 2026 expansion), plus $100B in committed AWS spend over ten years, plus a 5 GW Trainium compute allocation that effectively forces Anthropic onto AWS custom silicon at scale. Even on a $2T IPO, the multi-year AWS and Trainium revenue stream is contractually locked in. Amazon captures the per-token deflation as volume on its own infrastructure, with Trainium2 economics improving every quarter as utilization rises.

Alphabet sits on the other side. Its $200B cloud-spend commitment from Anthropic over five years is the single largest commercial TPU off-take in history — and because Broadcom co-designs the TPU silicon, that volume also flows through Broadcom's purchase-commitment backlog. The risk for Alphabet is that the same deflation pressure that's squeezing OpenAI and Anthropic also squeezes Gemini's enterprise pricing, but the upside is that the most credible non-NVIDIA AI compute stack in the world is being paid for by a customer that prints money at scale. Microsoft is the most exposed of the three hyperscalers: a 27% stake in OpenAI is now worth about $135B on paper, but Microsoft's economic claim on OpenAI is now capped and the April 2026 restructuring explicitly ended OpenAI's exclusive cloud arrangement.

Oracle is the most asymmetric name on the list. Its $300B Stargate deal with OpenAI —4.5 GW starting in 2027 — turns Oracle from a legacy database vendor into one of the top AI-cloud landlords in the U.S. The revenue is not yet in the run-rate (Q2 FY2026 revenue $71.8B), but Oracle's contract backlog disclosure includes Stargate as a material line item. Even if OpenAI's IPO slips again, the contract still ships. CoreWeave is the wildcard: a NVIDIA-backed pure-play with $22.4B in OpenAI commitments plus an Anthropic deal, but loss-making at the gross-profit line and exposed to the same token-price compression that is squeezing the labs.

The near-term catalyst calendar is unusually dense. Days–quarters: Anthropic's IPO roadshow is expected in October2026 against a $2T target — the Sept 22 price reset directly compresses the multiple underwriters can defend. OpenAI's listing slips into 2027, with the next inflection point being any Q4 2026 token-price announcement from either lab (most likely an Opus 5.5 promotional tier or a Fable 5.2). The first order impact is on the cloud-anchor P&L lines: Alphabet's Q3 2026 print will reflect the first month of Anthropic's expanded TPU off-take; Amazon's Q3 print will reflect the first month of the Trainium ramp; NVIDIA's Q3 FY2027 print (early November) will show whether the price war is reducing dollar-volume compute demand — early reads from the Blackwell Ultra cycle say no.

**1–3 years:</strong> the structural read is that per-token deflation will continue to compress lab gross margins while the dollar volume of compute keeps growing. The labs' IPO valuations will be forced to absorb that compression; expect either multiple compression from the $1T/$2T targets or revenue growth at multiples of current pace. The chip and cloud complex captures the upside regardless: NVIDIA and Broadcom sell into inelastic compute demand; TSMC fabricates both; Amazon and Alphabet host the workloads; Oracle and CoreWeave lease capacity into the same demand curve. The labs become the price-takers; the silicon vendors and cloud landlords become the price-setters. The next milestone to watch is Anthropic's October 2026 S-1 disclosure of forward gross-margin guidance — a number below 50% would be the cleanest signal that the Sept 22 price reset has permanently repriced the frontier.