Cheap Models Are Eating the Leaderboard—and the Bill Is the Feature

The most important number on this week’s model leaderboard is not 40.5 trillion. It is $0.10.

That is Anthropic’s price per million input tokens for Claude Haiku 5.5 below 100,000 tokens, and it helps explain why the small-model tier is turning into the execution layer for production AI systems. OpenRouter’s latest weekly usage table puts DeepSeek V4.1 Flash firmly in first place with 40.5 trillion tokens, StepFun’s Step 5 Preview in second with 15.1 trillion, and the newly released Haiku 5.5 at number 20 with 1.72 trillion. These models do not share a benchmark crown, a provider, or even the same product positioning. What they share is a credible case for doing a lot of work cheaply.

That matters because real agent systems do not make one heroic model call. They make dozens of mostly ordinary ones: classify the request, retrieve documents, compress context, choose a tool, normalize a response, check a field, and decide whether to continue. The expensive planner gets the demo. The cheap workers get the token bill.

Step 5’s jump is a deployment signal, not a quality verdict

Step 5 Preview added 3 trillion tokens between rolling snapshots, rising from 12.1 trillion to 15.1 trillion—a 24.8% increase—and passed GLM 5.3 Flash for second place. OpenRouter labels it “new,” meaning it had no usage in the preceding comparison window. That is a remarkably fast traffic ramp, especially for a model that still carries “Preview” in its name.

But OpenRouter measures prompt and completion tokens routed through its API. It excludes private requests, counts variants separately, and does not measure unique users, revenue, accuracy, intelligence, or successful tasks. A verbose model can generate more leaderboard credit than a concise one while completing the same amount of useful work. Step 5’s ascent therefore says that developers are sending it serious volume; it does not say those developers should trust it with an unreviewed database migration.

The sensible response is a controlled trial. Teams evaluating Step 5 should replay representative tool-using and long-horizon traces, then compare completed tasks, tool-call correctness, p50 and p95 latency, retries, and cost per accepted result. They should also retain a fallback and shadow a stable model against it. Preview is a lifecycle warning, not decorative branding: availability, behavior, and pricing can change before an architecture has paid back its integration cost.

DeepSeek V4.1 Flash provides the stronger proof that this pattern can persist. Its 40.5 trillion weekly tokens are 2.68 times Step 5’s volume, even after its displayed week-over-week growth cooled from 43% to 37%. DeepSeek has said third-party testing placed V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime, and it began routing V4 Pro API requests to Flash on September 14. That is more than a product launch: it is a provider using routing itself to make the faster model the default economic choice.

Haiku’s price is excellent—until the context window changes the arithmetic

Haiku 5.5 entered OpenRouter’s top 20 at number 20, displacing GLM 5.2 with 1.72 trillion weekly tokens. Anthropic is explicit about the intended jobs: summaries, context compaction, database queries, classification, customer support, browser use, and subagent work. Below 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5’s $1 and $5 rates.

The capability jump is substantial enough to make that price interesting. In Anthropic’s evaluation, Haiku 5.5 scored 39.2% on Terminal-Bench 4.0, compared with 0.0% for Haiku 4.5 and 16.4% for GPT-6 Luna. On the OSWorld 2.1 offline subset it reached 72.4%, versus 15.7% for Haiku 4.5 and 48.9% for Luna. Asana reported more than 30% lower task-completion latency and up to 2.5 times faster inference per agent turn in its own suite.

Those numbers do not make the flagship obsolete. Sonnet 5.5 scored 70.6% on Terminal-Bench and 83.9% on the OSWorld subset. The gap is the reason a tiered architecture works: a cheaper model can handle high-volume, bounded steps while a stronger model remains available for ambiguous planning, difficult coding, and recovery. Replacing every model with Haiku would miss the point. Replacing every easy call to Sonnet with Haiku, after measuring the failure rate, might materially improve the product.

There are two catches hidden behind the headline price. First, Haiku’s rates jump fivefold above 100,000 input tokens, to $0.50 per million input and $2.50 per million output. Simon Willison notes that GPT-6 Luna’s long-context increase begins at 272,000 tokens and is smaller. Second, his hands-on testing found that Haiku 5.5 used roughly 1.25 times as many tokens as Haiku 4.5 for the same long prompt. Token price is not task price when tokenizers, verbosity, retries, and context thresholds differ.

“Fast” is configuration-dependent too. Haiku 5.5 cannot disable reasoning and defaults to medium effort. In Willison’s SVG test, low effort took seven seconds and cost 0.0936 cents; maximum effort took five minutes and nine seconds and cost 3.3826 cents. Adjustable effort can be valuable, but only if the application sets it deliberately. Leaving effort at a generic default across classifiers, browser agents, and coding tasks is the AI equivalent of deploying every service with the same CPU limit.

Your router needs a cost model, not a model list

The practical work starts with traces. Divide production calls into narrow classes—classification, extraction, compaction, retrieval triage, tool selection, code generation, final synthesis—and build a small acceptance test for each. Then record quality, latency percentiles, input and output tokens, cache reads, retries, and escalations. Route each class to the cheapest configuration that clears its threshold, with a stronger model available when confidence or validation fails.

Do not optimize request cost in isolation. The right denominator is cost per completed task: a cheap model that needs three retries, emits 25% more tokens, or causes more human review can be the expensive choice. For support automation, measure resolved tickets and reopened cases. For coding agents, measure patches accepted after tests and review. For tool use, measure correct calls and recovery rate. A leaderboard can suggest candidates; it cannot define success for your system.

Context bands and caching also belong in the architecture, not in a procurement spreadsheet. Anthropic halved Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per million tokens and estimates roughly 20% lower cost for most agentic workloads. Agent loops repeatedly reread instructions, repository context, and prior tool results, so cache behavior can save more than a superficial model swap. Conversely, letting a Haiku prompt drift from 99,000 to 101,000 tokens changes its input rate by 5x. A dashboard that groups both calls under “Haiku traffic” is hiding the event that controls the bill.

Three moves are worth making now. Add context-band and cache-state dimensions to cost telemetry. Run Haiku 5.5 against compaction, classification, and subagent traces under 100,000 tokens, including low-effort settings. Put Step 5 Preview through a shadow evaluation on tool-heavy tasks, but keep a stable fallback until its behavior and lifecycle settle. None requires betting the product on this week’s ranking.

The leaderboard is beginning to resemble an infrastructure invoice rather than a beauty contest. DeepSeek’s lead, Step 5’s rapid climb, and Haiku’s arrival all suggest that production volume is flowing toward models that are fast and cheap enough to call repeatedly. The small-model tier is no longer the consolation prize. It is becoming the workforce—and the teams that measure total task economics, rather than repeating benchmark scores or token totals, will be the ones that actually capture the savings.

Sources: OpenRouter LLM Rankings, Anthropic: Introducing Claude Haiku 5.5, OpenRouter: Step 5 Preview, Simon Willison’s Haiku 5.5 hands-on, DeepSeek V4.1 Flash announcement