07 Mar 2024 AI Engineering 5 min read

Claude 3: A Portfolio Instead of a Single Model

Anthropic's Claude 3 family — Opus, Sonnet, and Haiku — replaces the single-model default with a tiered portfolio: one API, three price points, vision input, and a 200K context window across all tiers. We outline the verified launch numbers, the trade-offs, and a practical routing pattern that assigns each workload the cheapest model passing its evaluation.

The one-model default

Most production LLM systems in early 2024 run on a single model. One flagship handles everything: classification, extraction, summarization, code review, multi-step reasoning. The choice is made once, at project start, and rarely revisited. This is convenient. It is also expensive, because request distributions are skewed: the bulk of calls is simple, a minority is hard, and the two need different machinery.

A flagship model on a routing or tagging task wastes money and latency. A small model on complex analysis produces wrong answers. Neither failure mode becomes visible unless workloads are measured per task. With the Claude 3 family, announced on March 4, 2024, Anthropic makes the tiered alternative explicit: three models — Haiku, Sonnet, Opus — behind one API, at three price points.

A request arrives — the router classifies its difficulty. 1/4

The Claude 3 family in brief

Claude 3 comprises three models in ascending capability: Haiku, Sonnet, and Opus. All three share a 200K-token context window, accept image input, and carry a knowledge cutoff of August 2023. Anthropic states that all three can technically process inputs beyond one million tokens; that option is not generally available and is reserved for select customers.

At launch, Opus and Sonnet are generally available through the Anthropic API. Sonnet powers the free tier of claude.ai; Opus requires a Claude Pro subscription at 20 US dollars per month. Sonnet is also live on Amazon Bedrock and in private preview on Google Cloud Vertex AI. Haiku is announced but not yet released.

Capability and price per tier

The pricing spread is the central design decision. Opus costs 60 times as much as Haiku per token, on input and output alike. A team that sends every request to the top tier pays that factor on its entire volume, whether the task requires it or not. The published figures:

Benchmark figures come from Anthropic's model card and are indicative, not proof of fitness for a concrete application. Two details stand out. Opus reports 86.8% on MMLU (5-shot), narrowly above the published GPT-4 figure of 86.4%. And Haiku scores higher on HumanEval than Sonnet — 75.9% versus 73.0%. Small models are not uniformly worse; capability profiles differ by task.

Model	Input $/M tokens	Output $/M tokens	MMLU 5-shot	HumanEval 0-shot	Context
Claude 3 Opus	$15	$75	86.8%	84.9%	200K
Claude 3 Sonnet	$3	$15	79.0%	73.0%	200K
Claude 3 Haiku	$0.25	$1.25	75.2%	75.9%	200K

Vision input across all tiers

All three models accept images as input: photos, charts, technical diagrams, document pages. The capability is not reserved for the top tier. For document-heavy organizations this matters; Anthropic notes enterprise customers whose knowledge bases consist of up to 50% formats such as PDFs and presentation slides. On the AI2D science-diagram benchmark, Sonnet reports 89.2%, Opus 88.3%, and Haiku 80.6% (0-shot).

The scope is narrow: vision is input only. The models generate no images, and they return no reliable pixel coordinates or bounding boxes. Reading a chart works; object detection does not. A system that needs localization or segmentation still needs a dedicated vision model. Treat Claude 3 vision as document understanding, not as computer vision.

Long context and measured recall

A 200K-token window corresponds to roughly 150,000 English words. Long context is only useful if the model retrieves from it reliably. In Anthropic's Needle-in-a-Haystack evaluation — a planted sentence retrieved from a large corpus — Opus reaches 99.4% average recall and 98.3% at the full 200K length. Sonnet and Haiku land around 95%, ahead of Claude 2.1.

Two caveats. NIAH measures retrieval of one planted sentence, not reasoning across an entire window; a high score does not guarantee correct synthesis of 200K tokens. And long context is billed per token: a full 200K-token prompt to Opus costs about 3 US dollars in input alone, before any output. Long context replaces retrieval infrastructure only where request volume is low.

Choosing the model per task

The engineering consequence is a routing decision per workload, not per project. Classification, extraction, moderation, and simple transformations go to the smallest model that passes the task's evaluation — Haiku, once it ships. Retrieval-augmented answering and standard production traffic sit with Sonnet. Multi-step reasoning, difficult analysis, and low-volume high-stakes work justify Opus.

A second pattern is escalation: send the request to the cheap tier first, validate the output, and re-run on a higher tier when validation fails or confidence is low. Because all three models share one API shape, switching tiers is a configuration change, not a rewrite. The prerequisite is an evaluation set per task; without one, tier assignment is guesswork.

What tiering does not solve

Prompts do not transfer unchanged between tiers. A prompt tuned on Opus can degrade on Haiku, so each tier needs its own evaluation run — three models triple the test surface. Routing itself becomes a component that must be built, monitored, and versioned. And Haiku is announced, not shipped; production planning should not depend on it yet.

Cost does not fall automatically either. Cheap tokens invite volume, and total spend can rise while unit cost drops. Finally, tiering removes no model-level risk: Anthropic reports fewer unnecessary refusals and higher accuracy versus Claude 2.1, but hallucinations remain, on every tier. Output validation stays mandatory regardless of which model produced the output.

Outlook from March 2024

We expect tiered portfolios to become the norm across providers within the year. The pattern is too economically obvious to stay unique: OpenAI already spreads workloads across GPT-4 Turbo and GPT-3.5 Turbo, and Google positions Gemini in Ultra, Pro, and Nano tiers. The unit of model selection is shifting from provider to task.

Two predictions. First, mid-tier models will reach today's flagship quality within roughly a year, at a fraction of the price; teams that build routing now collect that gain automatically. Second, per-task evaluation sets will become the durable engineering asset — models get exchanged, evals persist. At Blue IT Systems we treat model choice as configuration, not architecture. Claude 3 is the first release that makes this stance the intended default.