← /editions
#16

Cheaper per token, pricier per task

// Anthropic launched Claude Sonnet 5, a mid-tier model that nears flagship Opus 4.8 on agentic benchmarks at roughly half the per-token price and is now the Free/Pro default — but an independent lab measured it at higher cost per finished task than Opus, spotlighting the tokenizer and effort level as the real cost levers.

$ today_in_10_seconds

Anthropic shipped Claude Sonnet 5, a mid-tier model that lands close to flagship Opus 4.8 on agentic benchmarks at roughly half the per-token price — and it's the new default on Free and Pro. But the sticker price hides a twist: an independent lab clocked it higher cost per finished task than Opus on standard pricing, because a new tokenizer and heavier reasoning make it emit more tokens. The real cost lever this release exposes isn't which model you pick — it's the effort dial most teams leave on autopilot.

$ lead_story

Anthropic launches Claude Sonnet 5, closing the gap to Opus 4.8

Anthropic released Claude Sonnet 5, pitched as its most agentic Sonnet yet — built to make plans, drive browsers and terminals, and run autonomously across long tasks, and now the default model on Free and Pro. On benchmarks it closes most of the distance to flagship Opus 4.8: 63.2% on SWE-bench Pro (vs Opus's 69.2% and Sonnet 4.6's 58.1%), 81.2% on OSWorld-Verified, and 80.4% on Terminal-Bench 2.1 — and it actually edges Opus 4.8 on the GDPval-AA v2 knowledge-work benchmark, 1,618 to 1,615. Pricing undercuts Opus clearly on paper: $2/$10 per million tokens through August 31, then $3/$15, against Opus 4.8's $5/$25. The agentic behavior is what early testers flagged: it finishes complex tasks where previous Sonnet models would stop short, and checks its own output without being asked. It's also safer than its predecessor — better at resisting prompt-injection hijacks, with lower hallucination and sycophancy rates.

// why_it_matters: Sonnet 5 stops being the budget fallback and becomes the rational default for most agentic workloads — which reframes model selection from 'which is smartest' to 'how much accuracy do I need per dollar.'

read more →

$ also_worth_attention

  • The catch: cheaper per token, sometimes pricier per task

    Here's the number almost no launch coverage led with — independent lab Artificial Analysis measured Sonnet 5 at about $2.29 per task versus roughly $1.99 for Opus 4.8 on their Intelligence Index at standard pricing — about 15% higher — because Sonnet 5 emits more tokens per task. Two things drive it: a new tokenizer (the same one from Opus 4.7) maps the same text to roughly 1.0–1.35× more tokens, and higher reasoning effort burns more. For anyone running agents at scale, the lesson is blunt: track cost per completed task on your own workload, not the per-token sticker.

    link →
  • Effort level is the cost dial you're ignoring

    The same analysis surfaced the release's most useful operational insight: max effort uses roughly six times the conversational turns of low effort on GDPval-AA — a six-fold swing that dwarfs the roughly two-fold per-token price difference between Sonnet 5 and Opus 4.8. In other words, effort level is the primary cost lever and model choice is secondary — most teams have it backwards, agonizing over routing while leaving effort at a reflexive default. And at its maxed-out xhigh setting, Sonnet 5 can cost more than Opus 4.8 at a comparable accuracy point, so dialing it blindly to max defeats the purpose.

    link →
  • The launch is also an IPO move

    The business framing is hard to miss: Anthropic released a near-flagship model at mid-tier prices as it races toward an IPO, putting powerful agentic capability in front of cost-conscious enterprise developers right as it courts public-market scrutiny. Making Sonnet 5 the free-tier default is a distribution play — it widens the top of the funnel while the introductory pricing makes switching close to cost-neutral. The read for founders: aggressive mid-tier pricing is becoming a strategic weapon in the model wars, not just a courtesy to developers.

    link →

$ bigger_picture

The trend behind the headlines

Sonnet 5 marks a shift in how the frontier is priced and chosen. For the first time Anthropic's top tier is a menu, not a flagship — Sonnet 5, Opus 4.8, and Fable 5 span a 5x price range, and they're meant to be mixed. The interesting consequence is that 'which model' is now the wrong first question. As benchmarks across tiers cluster within a few points, the variables that actually move your bill are the tokenizer, the effort setting, and how many steps a task takes to finish — none of which show up on a pricing page. This is the cost-side mirror of the 'harness over model' theme this newsletter keeps hitting: raw capability is converging and commoditizing, so the engineering edge (and the budget discipline) moves to how you route, tune, and instrument. The teams that win won't be the ones who picked the 'best' model; they'll be the ones measuring cost per finished task and treating effort as a first-class control.

$ tool_of_the_day

Artificial Analysis

The independent benchmarking lab whose per-task cost measurement is the reason today's edition isn't just a launch rewrite. It evaluates models on a standardized Intelligence Index and, crucially, reports cost per task rather than only per-token price — which is exactly the metric that flips the Sonnet-vs-Opus story. Before you set a production default off a launch chart, it's the reference worth checking against your own effort settings and workload. Better still: run your own eval, because token economics are workload-specific.

explore Artificial Analysis →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime