Model Releases Are Obsolete on Day One: Cost per Task Becomes the New Selection Benchmark

Grok 4.5, GPT-5.6, and Claude Sonnet 5 all shipped within a single week. A year ago, that density of launches would have triggered another round of "who's the strongest" leaderboard wars. In the 2026 reality, something more interesting is happening: a model can be obsolete on the day it ships — not because a stronger model beat it, but because a cheaper, better-matched calling pattern re-priced it.

The industry's main storyline has shifted from "can the model work?" to "can the model be continuously paid for at scale?" The numbers support the shift: Kimi's annual recurring revenue crossed the $100M, $200M, and $300M marks between March and June 2026; OpenRouter's weekly token volume grew 9.65% week over week; Tencent Hunyuan Hy3 grew daily token consumption 20x from preview to GA. Models are moving from low-frequency trials into persistent, high-volume workflows — and the selection yardstick is moving with them.

From "who's strongest" to cost per task

The old selection logic was simple: run the benchmarks, pick the highest score. But when most everyday tasks don't need frontier capability, "strongest" stops being a decision criterion. Cost per task is replacing per-token price as the metric that actually matters.

The rough formula:

cost per task = total model and tool spend ÷ number of outputs that clear the quality bar

That formula folds an entire task chain into the ledger: input and output token counts, number of call rounds, context length, tool-call frequency, reasoning effort, failure and retry rates, cache hit rate, and even the cost of human review. The competitive focus shifts from "maximize capability" to "optimize the cost of useful capability."

Why per-token price lies

Judging by "price per million tokens" alone produces systematic misjudgments. A cheap model can cost more per task if it needs more rounds, longer context, or higher retry rates. Conversely, a pricier model that gets the task right in fewer passes can end up cheaper overall.

Model routing quantifies this neatly. In a cascade, a cheap model tries first and a flagship redoes it if quality checks fail. Let c be the cheap model's cost, C the flagship's, and f the cheap model's failure rate — the cascade only pays off when f × C < (C − c). On the August 2026 price map, the spread between the cheapest and most expensive agent-capable models has widened to roughly 35x on output tokens — DeepSeek V4 Flash at ~$0.28 per million output tokens versus GPT-5.6 Sol at $30. "Just use the best model" at that spread isn't prudence; it's a budgeting decision made by accident.

Three vendor responses: tiering, effort, scenario-defined models

Faced with single-task economics, the major labs took three different engineering routes.

OpenAI went with scenario tiering. GPT-5.6 replaces the "flagship + mini" naming system with Sol / Terra / Luna. Sol handles complex reasoning and coding ($5/$30), Terra matches last generation's flagship at a mid-range price ($2/$15), and Luna targets high-concurrency, low-latency volume ($1/$6) — up to a 5x spread between tiers. The catch: the bare alias gpt-5.6 defaults to the most expensive Sol tier. If you don't explicitly specify a tier, production requests bill at flagship rates, which forces developers to think about tier assignment task by task.

Anthropic went with an effort mechanism. Claude Sonnet 5 drops traditional sampling parameters like temperature in favor of a reasoning-effort dial: low → medium → high → xhigh → max. Anthropic's measurements show medium effort on Opus 4.5 matches Sonnet 4.5's SWE-bench score while cutting output tokens 76%. "Which model" gains a companion question: "how much compute per call."

xAI went with scenario-defined models. Grok 4.5 was co-trained with Cursor, locking onto coding agents as its high-frequency scenario and using trillions of real interaction tokens to shape the model — priced aggressively at $2/$6. Musk's launch framing was blunt: the flagship is genuinely better, but most tasks don't need it.

Chinese vendors pushed cost-effectiveness into the architecture itself. DeepSeek V4 Flash has 284B total parameters but activates only 13B per token; Hunyuan Hy3 is 295B with 21B activated; Kimi K2.6 is 1T with 32B activated. The MoE "big parameters, small activation" route gives models a frontier-sized capability pool while keeping per-inference compute cheap.

The pricing calendar: selection is now a continuous task

"Obsolete on day one" has a more literal meaning: prices move faster than model releases. On July 30, 2026, OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 and Terra 20% to $2/$12. Claude Sonnet 5's introductory price of $2/$10 expired August 31 and rose to $3/$15 on September 1 — a 50% jump on the default tier overnight. Pricing pages run faster than blog posts, faster than your deployment. Locking onto one model is increasingly the most expensive option.

What teams should actually do: routing, swappable clients, in-house evals

The engineering playbook for this dynamic market is already taking shape. Model routing is the centerpiece: Amazon Bedrock's Intelligent Prompt Routing, Azure AI Foundry's Model Router, and ecosystem tools like OpenRouter, LiteLLM, and Dify all push "manually picking a model" toward "route each request by complexity, cost, and latency."

A swappable client is what makes routing real: every call goes through one interface, and the model is chosen by workload and configuration — never hardcoded at the call site. When the next price cut lands, moving traffic is a config edit, not a refactor.

Evaluation needs a new yardstick too. Cross-vendor benchmarks use different harnesses, so a two-point gap is essentially a tie. What matters is success rate on your own tasks, tool-call validity, and cost per completed task rather than cost per token. DeepSeek's first-party API offers roughly a 98% cache discount, and Kimi K3's cache-hit input drops from $3 to $0.30 — real ledgers you can only see with your own evaluation.

The industry impact: the default-call race and Jevons' moment

Single-task cost logic ultimately pushes competition toward a "default call" race: whoever lands in the default configuration of cloud platforms, AI gateways, and enterprise middleware captures the largest share of real traffic and real feedback, then feeds that into further cost reductions — a flywheel. Google embedding Gemini into Gmail and Docs, and Adobe embedding Firefly into creative workflows, are both plays for that entry point.

Deeper still, this points to the model world's "Jevons moment": when steam engines became more efficient, coal consumption went up, because lower usage thresholds expanded the demand frontier. The outcome of the price-performance war may not be lower total AI spend but higher AI usage density — model calls spreading from a few high-value tasks into everyday workloads. For enterprises, that means model selection is no longer a decision for one release season. It's an ongoing engineering capability measured in cost per task.