When Anthropic launched Claude Sonnet 5, the headline numbers looked generous: $3/$15 per million tokens (input/output) — identical to Sonnet 4.6, and well below Opus 4.8’s $5/$25. Introductory pricing cuts that further to $2/$10 until September 1, 2026. On paper, this is Opus-grade intelligence at Sonnet prices. Artificial Analysis ran the actual task economics and found something different.
According to Artificial Analysis’s Intelligence Index benchmark, Sonnet 5 costs $2.29 per task at max effort — roughly twice what Sonnet 4.6 costs, and approximately 15% more than Opus 4.8. The driver is not the token price. It is token volume. Sonnet 5 uses ~40% more output tokens per task than Sonnet 4.6 and generates roughly 3x more agentic turns on knowledge-work evaluations. It scores 53 on the AA Intelligence Index, a 6-point improvement over Sonnet 4.6 at max effort, placing it among the costliest models to run — behind only Fable 5.
This is not a rounding error. It is the core tension in every frontier model launch right now: the metric Anthropic controls (price per token) is not the metric that hits a developer’s invoice (cost per task completed).
Sonnet 5 priced at $3/$15 per million tokens — same as Sonnet 4.6, ~40% cheaper than Opus 4.8. Intro rate: $2/$10 until Sept 1, 2026.
Artificial Analysis Benchmark (Late June 2026)
Per-task cost measured at $2.29. ~40% more output tokens per task vs Sonnet 4.6. ~3x more agentic turns on knowledge-work evals.
The Gap Revealed
Sonnet 5 per-task cost exceeds Opus 4.8 by ~15%, despite cheaper per-token rate. Sits among most expensive models to run — behind only Fable 5.
The Structural Question (July 1, 2026)
Does “Opus-tier agents at Sonnet prices” hold when task cost exceeds Opus? The per-token price is a vanity metric. The P&L sees tasks.
The key insight: Per-token pricing is what Anthropic publishes. Per-task cost is what developers actually pay. When a model reasons harder and loops more, a cheaper rate card can produce a more expensive bill — and Sonnet 5’s ~3x turn expansion on agentic tasks is precisely that dynamic playing out at scale.
The Structural Read
This is a Product Overhang moment — but with an economic twist Anthropic didn’t advertise. The Product Overhang Doctrine says capability accumulates invisibly until it surfaces all at once. Sonnet 5’s capability jump is real: a 6-point Intelligence Index gain over Sonnet 4.6 at max effort is not noise. But capability in the agentic era does not arrive in isolation. It arrives with a token footprint.
Call it tokenmaxxing: more intelligent models, given latitude to reason, will use that latitude. They generate longer chains of thought, revisit their own outputs, and spin more agentic turns before arriving at an answer. That is the behavior you’re paying for when you step up from Sonnet 4.6 to Sonnet 5. The model isn’t wasteful — it is doing more work. But “doing more work” and “costing more money” are now the same sentence, regardless of what the rate card says.
This directly complicates the launch narrative that Sonnet 5 delivers Opus-tier agents at Sonnet prices. At the token rate, that claim is arithmetically true. At the task level — which is the only level that maps to a deployed product’s coststructure — the model costs more to run than Opus 4.8 itself. Developers building production agentic pipelines should model task cost, not token cost. The gap between those two numbers is now a competitive variable.
Product Overhang Doctrine — Applied
“Capability overhang surfaces all at once — but so does the cost overhang attached to it. Every intelligence gain that expands token usage is a hidden price increase wearing a cheaper rate card. In the agentic era, the model that reasons best is rarely the model that bills least.”
Three Implications
IMPLICATION 1 — DEVELOPERS MUST REMODEL COST STRUCTURES
Any team that built budget forecasts on Sonnet 5’s token pricing alone is carrying a gap in their unit economics. Artificial Analysis’s data — $2.29 per task, ~40% more output tokens, ~3x agentic turns — needs to be the input to pipeline cost modeling, not the rate card. This is especially material for high-volume knowledge-work automations where agentic turn counts compound quickly.
IMPLICATION 2 — THE FRONTIER PRICING WAR IS BEING FOUGHT ON THE WRONG METRIC
Per-token price wars — OpenAI, Anthropic, Google all cutting input/output rates — are increasingly disconnected from what enterprise buyers care about: cost per outcome. As models get smarter and more agentic, per-token pricing becomes a marketing lever, not an economic signal. The vendors who win enterprise contracts will be those who can guarantee predictable per-task cost, not just cheap tokens.
IMPLICATION 3 — ANTHROPIC’S SONNET/OPUS TIER DISTINCTION IS BLURRING
When Sonnet 5’s per-task cost exceeds Opus 4.8 by ~15%, the tier separation that justified keeping two products in the lineup weakens. Anthropic retains the token-price differential ($3/$15 vs $5/$25), but the operational cost advantage has flipped. That creates a confusing purchase decision for developers: do you choose Sonnet 5 for its Intelligence Index score of 53, or Opus 4.8 for potentially lower real-world cost on tasks that don’t require maximum reasoning depth?
Business Engineer Framework
Product Overhang Doctrine
The Sonnet 5 story is a textbook Product Overhang case — capability accumulates invisibly, then surfaces with unexpected economic consequences. The Map of AI maps where each model sits across the 9 layers of the AI stack, including the inference economics layer where per-task cost lives. Understanding which layer you’re competing in changes how you read every model launch.
Anthropic gave developers a smarter model at a cheaper token rate and called it a deal — but Artificial Analysis’s numbers expose the fine print: Sonnet 5 costs $2.29 per task, ~15% more than Opus 4.8 and roughly twice Sonnet 4.6, because intelligence in the agentic era is paid for in tokens consumed, not tokens priced. Until the industry standardizes on per-task cost as the primary benchmark, every model launch that leads with token pricing is telling you the number that looks best, not the number that matters.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
