Anthropic Faces a Question: How Is Sonnet 5 Costing More Than Fable 5?

By Kexi, Aofei Temple QuantumBit | QbitAI

Claude has received yet another update.

This time, however, is not a flagship release. Anthropic has introduced Sonnet 5, a new cost-efficient model.

Positioned as the 'most capable' Sonnet to date, the model can independently plan tasks, invoke browsers, and operate terminals.

Its benchmark scores approach those of Anthropic's premium Opus 4.8, yet at roughly 60% of the price, it functions as a de facto 'Opus alternative.' [IMAGE:0]

The figures are clear: Sonnet 5 achieved an agentic coding score of 63.2% on SWE-bench Pro, surpassing Sonnet 4.6 by 5 percentage points.

The listed price remains identical to that of Sonnet 4.6. Per the launch messaging, capabilities have improved while pricing has not.

Did pricing truly remain unchanged?

Developer Simon Willison performed a simple test: he fed identical text to both the old and new tokenizers and compared the token counts.

The finding: despite an unchanged nominal price, token consumption per task has stealthily increased by 30%.

Anthropic appears to be executing a classic bait-and-switch.

'Opus Alternative'

The key upgrade in Sonnet 5 centers on enhanced agentic capabilities.

The model can autonomously decompose tasks, invoke tools such as browsers and terminals, execute multi-step workflows end-to-end without interruption, and self-verify its output without human prompting.

The benchmark results show clear incremental gains.

On SWE-bench Pro, the agentic coding benchmark, Sonnet 5 scored 63.2, versus 58.1 for Sonnet 4.6 and 69.2 for Opus 4.8 — placing it between generations and within 6 points of Opus.

On OSWorld-Verified, the computer use benchmark, Sonnet 5 achieved 81.2% against Opus 4.8's 83.4%, narrowing the gap to 2.2 percentage points.

On GDPval-AA v2, a knowledge work benchmark, Sonnet 5 scored 1,618 — 3 points above Opus 4.8's 1,615. [IMAGE:1]

Early feedback from two companies that deployed the model corroborates this assessment.

Zimu Li, an engineer at AI coding platform Factory, noted that Sonnet 5 provides their agents with robust execution capabilities, enabling sustained coding, tool invocation, and troubleshooting in complex technical environments — particularly well-suited for workflows demanding extended engagement and meticulous technical attention.

Daniel Shepard, an engineer at automation platform Zapier, provided a more concrete illustration. Sonnet 5 was assigned a two-stage task: updating customer account tiers in Salesforce, followed by sending a product launch announcement to enterprise clients.

Such tasks previously stalled midway — account tiers updated but announcements unsent, or the reverse. Sonnet 5 completed both stages end-to-end without requiring human intervention.

In Shepard's assessment, for routine automation, this model is the obvious choice. [IMAGE:2]

The safety evaluation results concurrently released by Anthropic align with this overarching thesis.

Sonnet 5 exhibits lower hallucination rates and reduced sycophancy compared to Sonnet 4.6, and demonstrates greater resilience against prompt injection and similar hijacking attacks in autonomous tool-use scenarios. [IMAGE:3]

The significance of these performance metrics only emerges when weighed against pricing.

Opus 4.8 is listed at $5/$25 per million input/output tokens, while Sonnet 5 is priced at $3/$15 — approximately 60% of Opus, or 40% with the limited-time promotion through end of August.

On paper — and this qualifier warrants attention — performance exceeds 90% of Opus at 40-60% of the cost, positioning Sonnet 5 as a de facto Opus alternative.

Thus concludes the performance narrative. The pricing story now reveals its other dimension.

Sonnet 5 employs a new tokenizer — the mechanism by which the model segments text into tokens.

Under the new tokenizer, identical text is now segmented into a greater number of tokens, inevitably driving up the total cost.

The discovery was made by developer Simon Willison.

Rather than accepting the official 'price unchanged' claim, he used his own token-counting tool to test several real-world documents across different languages.

In English, Sonnet 4.6 produced 2,356 tokens versus Sonnet 5's 3,341 — a 42% increase. Spanish saw a 33% rise.

A 4,279-line Python codebase saw token counts rise from 44,014 to 56,118, a 27% increase.

Simplified Chinese was the sole outlier, rising from 3,334 to 3,360 — a mere 1% increase.

The same tokenizer upgrade yields impact disparities exceeding 40 percentage points across languages — a discrepancy more noteworthy than the price increase itself. [IMAGE:4]

Taken together, these figures show that the 'unit price unchanged' claim is technically accurate yet materially incomplete.

An identical English passage now costs more than 40% above its Sonnet 4.6 equivalent.

This, however, should still leave it cheaper than Opus — until a developer promptly demonstrated otherwise.

The developer discovered that, in practice, Sonnet 5 costs more than Opus.

On the Artificial Analysis Intelligent Index, Opus 4.8 averaged $1.80 (weighted) per task, versus Sonnet 5's $2.29 — a 27% premium. [IMAGE:5]

He further found that, for identical tasks, Sonnet 5 consumes up to twice the tokens of Opus. [IMAGE:6]

Across the full benchmark run, Sonnet 5 costs 6.8% more than Fable. [IMAGE:7]

Fortunately, token-counting tools enabled developers to precisely quantify this discrepancy — which is why Willison was able to surface the 'stealth price increase' on launch night itself.

Ordinary users would likely only notice a vague sense that 'the bill seems higher this month' — precisely the stealth mechanism Anthropic employed.

How should developers navigate this landscape?

Placing Sonnet 5's 'stealth price increase' in industry context, price hikes themselves have been common over the past two years. The real distinction lies in the approach taken.

OpenAI's April price adjustment for GPT-5.5 followed the transparent route: per-million input/output token pricing was raised from $2.5/$15 to $5/$30 — a full doubling, visible to anyone visiting the pricing page.

Sonnet 5 took a different approach: the listed price remained untouched, with the increase concealed within the technical detail of the tokenizer swap — undetectable without active measurement.

For developers migrating workflows to Sonnet 5, one exercise is more valuable than tracking 'whether the price card changed': measuring actual token consumption against real workloads to estimate effective cost.

The price card only reveals the unit rate; the bill alone discloses the true cost of this upgrade.

The quietly rewritten invoice reveals that 'Opus alternative' is merely two sides of the same coin.

Sonnet 5 indeed delivers Opus-approaching capabilities at a lower price — except that the phrase 'lower price' now conceals a recalibrated yardstick.

Finally, one developer noted that Claude is not the only available option.

The developer observed that while Sonnet represents a cost-effective option within Anthropic's own lineup, Zhipu GLM-5.2 delivers performance comparable to Sonnet 5.

In terms of pricing, GLM-5.2's input cost stands at just 70% of Sonnet 5's, with output cost below half. [IMAGE:8]

In the developer's own words: why spend more to get less?

Reference links: [1]https://www.anthropic.com/news/claude-sonnet-5 [2]https://simonwillison.net/2026/Jun/30/claude-sonnet-5/ [3]https://x.com/theo/status/2072068395529576912 [4]https://x.com/0xSero/status/2072028529550832068