Claude 3 in three tiers: Haiku, Sonnet and Opus

Anthropic announced the Claude 3 model family on March 4, 2024, split into three tiers in ascending order of capability: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. Opus and Sonnet were available at launch in claude.ai and through the Claude API, which Anthropic simultaneously moved to general availability in 159 countries; Haiku was slated to follow shortly after.

Three tiers, three price points

The tier split is the practical heart of the announcement. All three models ship with a 200K context window at launch, and the pricing per million tokens (input / output) is:

Claude 3 Opus: $15 / $75
Claude 3 Sonnet: $3 / $15
Claude 3 Haiku: $0.25 / $1.25

Anthropic positions the tiers by workload: Opus for highly complex, open-ended tasks (research review, advanced chart and market analysis, interactive coding), Sonnet for high-volume enterprise workloads (RAG and retrieval over large knowledge bases, sales automation, code generation), and Haiku for near-instant responses (live support, translation, content moderation). Anthropic says all three models also accept inputs beyond 1 million tokens for select customers who need the extra headroom.

The benchmark claims

Anthropic says Opus outperforms competing models on most common evaluation benchmarks, including MMLU (undergraduate-level expert knowledge), GPQA (graduate-level reasoning) and GSM8K (basic mathematics), and that the family as a whole improved on non-English conversation, code generation, and analysis and forecasting.

On accuracy, the company tested the models against a set of complex factual questions that target known model weaknesses, sorting responses into correct, incorrect (hallucination) or "I don't know". By Anthropic's numbers, Opus roughly doubled the correct-answer rate versus Claude 2.1 while also reducing wrong answers. The announcement's footnote is worth keeping in mind: engineers who run these evaluations have optimized prompts and few-shot samples, and reported higher scores for a newer GPT-4T model.

Speed, vision and recall

Three capability areas stand out for production use:

Speed. Anthropic claims Haiku can read an information-dense research paper of roughly 10,000 tokens, including charts and graphs, in under three seconds, and that Sonnet is 2x faster than Claude 2 and Claude 2.1 for the vast majority of workloads while being more capable.
Vision. All Claude 3 models process photos, charts, graphs and technical diagrams. Anthropic notes that some of its enterprise customers have up to 50% of their knowledge encoded in formats like PDFs, flowcharts and presentation slides.
Long-context recall. On the "needle in a haystack" test, a measure of recalling buried facts from a large corpus, Anthropic reports Opus above 99% accuracy, and in some runs the model even flagged that the "needle" sentence looked artificially inserted into the text.

Fewer refusals, a safety rating

Anthropic says Claude 3 models are significantly less likely to refuse harmless prompts that border on the guardrails, and that bias scores improved on the BBQ benchmark versus earlier models. The family remains at AI Safety Level 2 (ASL-2) under Anthropic's Responsible Scaling Policy, and the company's red teaming concluded the models present negligible potential for catastrophic risk at this time, while staying below the ASL-3 threshold. Citations pointing to specific sentences in reference material were promised to arrive soon.

What to take from this

If you're picking a model for a workload, the criteria are now explicit: latency-sensitive, high-volume, simple tasks point to Haiku at a fraction of the cost of a frontier call; general production workloads point to Sonnet, where the price-to-capability balance is the strongest; Opus is for the problems that actually need the most intelligence, at a price you'll want to meter. Anthropic also announced that Tool Use (function calling), interactive coding (REPL) and more advanced agentic capabilities were coming for the family in the months ahead, worth watching if your stack depends on structured output and tool calls.