Claude Sonnet 5: Behind the Price Cut, Per-Task Costs Actually Rise

---

TL;DR

• Sonnet 5 lowers list prices while claiming performance near Opus 4.8, but Artificial Analysis’s measurements show per-task cost is roughly double that of Sonnet 4.6, and about 15% higher than Opus 4.8. The $/token headline price is no longer a reliable guide.
• Sakana Fugu handed its orchestration layer to Google Cloud, and Meta has started reselling its surplus compute — the scramble for infrastructure is shifting from “build” to “sell.”
• Sarashina3, Gemini Spark, and the Google Home Speaker all show the same pattern: capable models paired with implementation/rollout that’s still catching up.

Top Pick: Claude Sonnet 5 — Behind the Price-Cut Headline, Per-Task Costs Rise

On June 30, Anthropic announced Sonnet 5 and made it available immediately across all plans and the API. The company claims it delivers performance approaching flagship model Opus 4.8 at a lower price than Opus 4.8 ($3 input / $15 output per million tokens, with a limited-time promotional rate of $2 input / $10 output through August 31). But measurements from third-party evaluator Artificial Analysis expose what’s really behind this “cheaper yet Opus-equivalent” narrative.

The Technical Read

Raising the Effort (reasoning depth) level lets Sonnet 5 match Opus 4.8 on some tasks, such as agentic search. Its composite Intelligence Index score climbed from Sonnet 4.6’s 47 to 53, closing the gap to just 3 points behind Opus 4.8 and 2 points behind GPT-5.5. Gains in tool use, coding, and knowledge work are real, but a good share of this progress looks less like pure model improvement and more like “buying performance by turning up the inference-time billing dial called Effort.” Safety improved overall versus Sonnet 4.6 (better refusal of harmful requests, stronger resistance to prompt injection, fewer hallucinations), but the success rate of vulnerability exploitation actually rose compared with Sonnet 4.6 — hence the caveat that Anthropic is shipping it with detection/blocking features enabled for dangerous use cases.

The Business Read

Looking only at the headline numbers, it seems like “Opus-level performance at 60% of the price.” But Artificial Analysis points out the catch: Sonnet 5 consumes roughly 40% more output tokens per task than Sonnet 4.6, so the actual cost per task swells to roughly double. It’s also about 15% higher than Opus 4.8 on a per-task basis. In other words, under a design where raising Effort increases output token consumption, the headline ”$/M tokens” price no longer functions as a reliable predictor of your actual bill. The implicit assumption that “a new model generation equals lower cost” no longer holds.

Implications and Positioning

• What to build or drop: Auto-upgrading to Sonnet 5 is a performance upgrade, not a cost cut. For high-frequency, low-budget API workloads, sticking with the prior generation model or a lower Effort setting may end up cheaper.
• Hype to ignore: Don’t take the “it’s cheaper, so you should switch” narrative at face value.
• Stance update: Shift your model-selection criteria from the ”$/M token” headline price to the “measured per-task cost on your own primary workloads.” Auto-updating your default model without measuring impact is a quiet cost-inflation risk going forward.

Signal/Noise call: Signal (high confidence). This is based on measured data from a third-party benchmark and is consistent with Anthropic’s own safety report (rising vulnerability-exploitation success rate → added blocking features).

Other Key Topics

Sakana Fugu Builds on Google Cloud — Multi-Agent “Plumbing” Gets Absorbed by the Platform

It’s now known that Sakana AI’s multi-model routing system, Fugu/Fugu Ultra, has adopted Google Cloud’s Gemini Enterprise Agent Platform (formerly Vertex AI), Cloud Run, Model Armor, Firebase, and BigQuery across every layer, from the frontend down to the model layer. Technically, there’s nothing novel about orchestration that switches between multiple models based on task difficulty, but Fugu Ultra claims to outperform Claude Mythos Preview and Claude Fable 5 on some tasks — the source of its strength lies not in the models themselves but in the routing design. That’s the key business point: Sakana AI itself explains this as “a natural move given our roots in ex-Google/DeepMind talent,” but the real reason for the choice is infrastructure stability and managed multi-model integration features. In other words, this “smart routing” is a textbook example of logic that could largely become commoditized once Google natively strengthens orchestration features within the Gemini Enterprise Agent Platform. The only asset that could be called proprietary is the accuracy of deciding “which model to assign to which task” — a thin moat. Implications and positioning: If you’re building your own multi-model router/orchestration layer, base your investment decisions on the assumption that Google/AWS/Azure will eventually offer the same capability as a managed service. Put your differentiation not in infrastructure but in proprietary evaluation data that improves task-to-model assignment accuracy.

Meta Moves to Resell Surplus Compute as a Cloud Business — The Infrastructure Power Struggle Enters Its Next Phase

According to a Bloomberg report, Meta is preparing a cloud business to sell external access to its AI compute and models, positioning itself to compete with AWS, Google Cloud, and Azure. This follows SpaceX (via xAI) having Anthropic buy compute capacity from its “Colossus 1” data center, and subsequent similar lease agreements with Google and Reflection AI. Meta had already committed $182.9 billion to AI infrastructure by the end of Q1, with large-scale projects in Louisiana and Ohio also underway. The fact that compute stockpiled for internal demand is simultaneously becoming “sellable inventory” for both the Meta and SpaceX camps supports the idea that the winners may end up being compute owners rather than model providers. There’s skepticism about infrastructure expansion that depends on rapidly depreciating chips, but a contrarian read is also possible: monetization pressure is pushing compute accumulated beyond internal demand out into external sales. Implications and positioning: A growing number of compute sellers is a medium-term downward pressure on GPU/cloud pricing. There’s little reason to lock in large compute commitment contracts right now — bet on prices loosening within six months to a year.

Sarashina3: A Domestic LLM Trained via Self-Distillation (OPSD) Without Relying on External Models

SB Intuitions (a SoftBank subsidiary) has begun offering the Sarashina3 series (mini/nano/guard/embedding/rerank). It was pretrained on over 30 trillion tokens of data, mostly Japanese and English, and its technical distinguishing feature is that post-training combines reinforcement learning with “OPSD” (on-policy self-distillation) — using the model’s own outputs for training rather than “distillation” that borrows knowledge from an external high-performance model. This design, which cuts dependence on external models, strengthens independence in terms of licensing and data governance. On benchmarks it sits in the same tier as Qwen3-235B-A22B-Instruct-2507 and GPT-5.4 mini (non-reasoning) — mid-range rather than frontier-class. From a business standpoint, this is clearly positioned to target demand from government agencies and domestically regulated industries seeking sovereignty (domestic data centers, domestic data), so its TAM is limited not to the global general-purpose LLM market but to “domestic projects where data sovereignty is a contractual requirement.” Implications and positioning: Irrelevant to you as long as you need frontier-level capability. But if you handle projects — Japanese government, finance, healthcare — where contract requirements specify domestic data centers or domestic models, Sarashina3 emerges as an option for use cases where mid-range performance suffices.

Gemini Spark Comes to Mac — Desktop AI Agents Become a Battleground Where Differentiation Disappears

Google’s agent assistant Gemini Spark has been integrated into the macOS Gemini desktop app, adding Google Tasks/Keep integration and real-time information tracking (currently a beta limited to Google AI Ultra subscribers in the US). This marks a direct entry into the same “desktop agent that can touch local files” market as Claude Desktop, Copilot, and OpenClaw. Technically there’s little that’s new — this is more a matter of catching up with the feature sets of existing rival desktop agents (with file-to-document generation, like building a budget spreadsheet from invoices, as the main use case). From a business standpoint, this field is a textbook example of vendors converging on the same feature set (file manipulation, document generation, task integration), where niche desktop AI tools sitting alongside agents bundled natively into general-purpose OSes get dissolved by price and feature competition. Implications and positioning: If you’re building or considering a persistent desktop AI agent tool, redesign on the assumption that “file manipulation” and “task integration” features won’t differentiate you. The path to winning is deep integration into a specific workflow, or a specialized domain that OS vendors deprioritize.

Google Home Speaker: Great Hardware, Gemini Not Ready Yet — A Case Study in the Capability-Implementation Gap

In a review from The Verge, Google’s first new smart speaker in six years, the Google Home Speaker, earned high marks for the hardware itself, but the “Gemini for Home” software running on it was called “unfinished.” This is a case where even Google, the biggest platform player of all, hasn’t caught up on quality when integrating its latest model into its latest hardware. There’s a gap — one that even a deep-pocketed giant can’t close — between the expectations set by demos of “an agent that handles every household task” and an implementation robust enough for daily use. Implications and positioning: The bottleneck isn’t model capability but reliability in always-on environments. If you’re allocating budget to voice agent integration now, put it toward reliability measurement and fail-safe design rather than model selection.

Try This Week / Hype to Ignore

Try this week

• Before switching your default model, measure actual token consumption and cost for Sonnet 4.6 versus Sonnet 5 on your own primary tasks (coding, research, summarization, etc.) under identical conditions (fixed Effort level).
• If you’re building multi-model routing in-house, take inventory of how much proprietary data is going into the decision logic (task-to-model assignment accuracy). Keep additional investment in the plumbing itself to a minimum.

Hype to ignore

• The summary that “Sonnet 5 is a low-cost model matching Opus 4.8” — that’s the headline price, not the actual cost.
• “Fugu Ultra beats Mythos” — a claim limited to specific tasks, with a thin defensible moat in the routing layer itself.
• The Gemini Spark Mac support news — a lockstep advance with almost no feature differentiation from competitors.