The slogan is “racing to the bottom” — but to which bottom? In recent weeks Anthropic Opus 5.5 and OpenAI’s GPT-6 family (Soul, Luna, Astra) cut prices in parallel: Opus 5.5 from $5/$25 to $4/$20 per 1M input/output (−20%) , cache reads from $0.50 to $0.20 (−60%, a 95% discount vs. uncached) ; GPT-6 Soul undercuts its predecessor GPT-5.6 Sol by 50% ($2/$10) . On paper intelligence is getting cheaper. The video’s first objection starts there: a cheaper price tag does not mean a cheaper model.
Two Different Paretos: Cost Efficiency vs Token Efficiency
The frontier we usually track is Intelligence vs Cost per Task — the dashed line marks the ideal trade-off and we celebrate when a model pushes it outward. On this axis four of Opus 5.5’s five effort levels (max, xhigh, high, medium) sit on the Pareto frontier , and at 58 on the Artificial Analysis Intelligence Index it is the highest score measured to date, either cheaper or smarter than GPT-6 Astra, Fable 5.1 and Opus 5 scoring 50+. Concretely: Opus 5.5 (max) averages $0.088 / 1,434 tokens per task vs Opus 5’s $0.173 / 2,227 (−36% tokens, half the cost) ; GPT-6 Soul averages $0.118 / 2,178 vs GPT-5.6 Sol’s $0.225 / 2,521 (−12% tokens, half the price) . Much of Soul’s saving comes from the price cut itself.
Flip the same models to Intelligence vs Tokens and the ranking inverts. For a single Intelligence Index task Opus 5.5 (max) burns ~119k output tokens , versus ~73k for Opus 5, ~78k for Fable 5.1 and just ~27k for GPT-6 Astra . The GPT-6 family — Astra included — delivers the same intelligence far more taciturnly; Opus 5.5 only compensates by pushing the ceiling higher, plateauing around 53 on the index before token efficiency rolls over. The video stresses that GPT-6 holds a sustained token-efficiency advantage until Opus outscales it on sheer intelligence.
In 3D the True Price Appears
Hence the video lifts the chart into 3-D: intelligence – cost – tokens . Opus 5.5’s trajectory plateaus near 53 and simultaneously drifts away from the baseline toward token inefficiency; add an idealized baseline and the gap becomes obvious. Prior Anthropic models (Fable 5.1, Opus 5) show the opposite bias — more disciplined on tokens, pricier on cost. GPT-6 Soul (~47.5) and Luna (~37.3) trace a completely different path at a lower intelligence ceiling; Luna is strikingly cost-efficient when viewed irrespective of ceiling, while Opus 5.5 hugs the baseline on both cost and tokens only to fall off on tokens, and Astra’s path resembles Fable 5.1 more than Opus 5.5.
Subscription vs API: Who Pays for Verbose Tokens?
A sponsored interlude — Hyper Agent demoing a four-agent Italy trip (Sofia the planner, Marco on flights, Gianni on visuals, with cascading delegation) — brackets the core economic split: OpenAI is subscription-heavy, Anthropic API-heavy in revenue through 2025. The unit differs: subscriptions are metered by total tokens in a 5-hour window , APIs by dollars per token . When a model is verbose, API builders pay immediately; subscribers feel it less, depending on how permissive the lab is. Labs subsidize token usage at their discretion — competition, data-center availability, hardware cost, margins. That tension — who bears the cost of inefficient tokens — makes a token-efficient model a hedge against the relentless fall in the price of intelligence. Verbose tokens become the provider’s problem or the user’s, depending on the business model.
Not Just Anthropic vs OpenAI: Everyone Races to Cheaper
Zoom out and the trajectory is crowded: Moonshot’s Kimi, Z.AI’s GLM, DeepSeek, Xiaomi’s MiMo, Google’s Gemini, MiniMax M series, xAI and Meta all push costs down, with DeepSeek notably aggressive on the chart. Looking at a single model release therefore misleads; lower price of intelligence does not automatically mean intrinsically more efficient models . The video tracks the Pareto frontier quarter by quarter in 2026: from Q1 through Q3 both the intelligence-cost and intelligence-token frontiers inch toward the ideal , but not yet as a frontier that expands cleanly on both axes at once. The closing hint is to add the next dimension — speed — which would rewrite how we judge progress: the same intelligence, delivered faster, is a different kind of efficiency.
Key moments
- Racing to the bottom: does cheaper price mean cheaper intelligence?
Opus 5.5 −20%, GPT-6 −50% — but sticker price ≠ cost per task
- Two Paretos: cost vs token efficiency
Opus 5.5 on cost frontier yet 119k tokens vs Astra’s 27k — most frugal vs most verbose
- 3-D view: intelligence-cost-tokens together
Opus 5.5 plateaus near 53, drifting off baseline; gap visible once ideal line is added
- Subscription vs API: who bears verbose cost?
5-hour token window vs $/token — Hyper Agent demo frames the split
- Everyone races down: DeepSeek most aggressive
Kimi, GLM, DeepSeek, MiMo, Gemini, MiniMax, xAI, Meta — Q1→Q3 Pareto inches toward ideal, speed next
AI commentary
"What struck me is that the video compares not two prices but two economies: money out of your wallet and tokens burned by the model. Praising Opus 5.5 on price tag alone ignores how chatty it is — for me the lesson is that cheaper intelligence alone isn’t enough."
AI assessment
In my view the video’s strongest move is lifting a one-dimensional price chart into 3-D — suddenly you see which axis each celebrated “cheaper” model borrows from; both Opus 5.5 and GPT-6 look like a win and a trade-off at once. In my experience that matters most for teams running coding agents over the API, where the bill inflates with chattiness, not sticker price. What’s missing is precisely what the narration only whispers: speed as a fourth axis. A model delivering the same intelligence in half the time has economic value comparable to token frugality, yet it goes unmeasured here.
To steelman the other side: someone arguing “only the subscription price matters to the end user” could dismiss this 3-D rigor as pedantic — if ChatGPT Plus is $20 flat, token counts feel theoretical. They have a point, but only inside the subscription world. For anyone shipping a product on the API, chaining agents, or bumping into rate limits and 5-hour quotas, the argument collapses; inefficient tokens turn into latency and retries. The video earns its keep by making that split explicit, but I wish the minutes spent on the Hyper Agent Italy demo had been spent benchmarking subscription vs API load on a real agentic workflow — then “who pays?” would be a log line, not speculation.
My takeaway: cheer cheaper intelligence, but don’t worship a single metric. I’d reach for Opus 5.5 on ceiling tasks where peak intelligence justifies chatter, and for Luna/Astra on high-volume, low-margin work where cost and speed dominate. At decision time I would re-check the 53/47.5/37.3 ceilings and the $4/$20 vs $2/$10 tags against the live price page — in this race stickers change monthly, Pareto moves quarterly.
Sources
5 links; 1 of them also cited by 1 other story. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
- @youtube.com youtube.com · watch
- @artificialanalysis.ai https://artificialanalysis.ai/articles/claude-opus-5-5
- @claudecodecamp.com https://www.claudecodecamp.com/p/why-new-models-get-cheaper
- @anthropic.com https://www.anthropic.com/pricing
Also cited by: Make Money with Claude Cowork: The 2026 Full Course, Decoded
- @openai.com https://openai.com/pricing
opus 5.5 · gpt-6 · pareto efficiency