Claude

Why this family matters

Claude 4 is Anthropic's current generation: Opus (flagship), Sonnet (workhorse), Haiku (cost tier), and the Mythos preview at the top of the roadmap. The structurally interesting pattern across the family is the Opus pricing reset between 4.0/4.1 and 4.5: input dropped from $15 to $5 per million, output from $75 to $25. That is a 3x cost cut on the same brand-name SKU, and it materially changes the "should we even consider Opus?" conversation for workloads that were previously priced out.

The Sonnet 4.5 line ships a 1M-token context window at the same headline

input /

output as the 200K variants. As with the Opus reset, "long context" stops being a premium SKU and becomes a free axis on the parts of the family where it ships.

Which variant to start with

Default to anthropic-claude-sonnet-4. At $3 input / $15 output per million it is the family's price-quality sweet spot, and the 4.6-thinking variant lands at Quality Score 96.7 (#16 of 186 models we track), inside the cluster where the closed-flagship reasoning conversation actually lives. For most teams shipping API-backed product features, this is the practical default.

Step up to anthropic-claude-opus-4 when you can name the workload that justifies the cost gap. With the 4.5+ reset, the Opus premium over Sonnet is roughly $5 vs $3 input and $25 vs $15 output per million. That is far narrower than the pre-reset gap, but still material at production volume. Reach for Opus when agentic coding (SWE-Bench Verified) or hardest-tier reasoning (GPQA Diamond, AIME) is the binding constraint; 4.7-thinking is rank 2 / rank 4 / rank 3 on those respective benchmarks in our index, which is the ceiling signature you are paying for.

When to deviate:

• Coding agents: use anthropic-claude-opus-4 (4.5-thinking or newer). Opus 4.7-thinking is rank #3 of 68 on SWE-Bench Verified in our index, and the price reset makes the cost gap to Sonnet defensible for agentic-coding loops where reliability compounds.
• High-volume chat at scale: drop to anthropic-claude-haiku-45 ($1 / $5 per million). The score gap to Sonnet is real (Haiku 4.5 thinking at QS 77.9 vs Sonnet 4.6 thinking at QS 96.7), but for repetitive low-stakes turns the per-token cost cut dominates the unit economics.
• Long-context RAG: use anthropic-claude-sonnet-4 (4.5 variants). The 1M context window at Sonnet pricing is the cheap long-context play in the family. Opus 4.6-thinking also reaches 1M context but at the higher Opus price point; Sonnet 4.5-thinking at the same window is the procurement-friendly choice unless you have measured an Opus-specific quality gain on your eval.
• You are tracking the frontier: the Mythos preview lands at Quality Score 118.9 (#1 of 186) with first-place finishes on GPQA Diamond and SWE-Bench Verified in the data we have. It has no public pricing or general availability, so it is a roadmap signal, not an option to ship today. Treat its scores as Anthropic's benchmark publication, not as independently verified.

Where the data is weak

We aggregate benchmark scores from multiple sources but coverage and naming across this family deserve a careful read. Specifically:

• Opus has many minor versions (4.0, 4.1, 4.5, 4.6, 4.7, plus Thinking modes for each). Within-family scores vary substantially across these (e.g. Opus 4.1 non-thinking at QS 70.4 vs Opus 4.7-thinking at QS 107.8). When the article quotes a number, it is for the specific minor version named; do not collapse the line to a single Opus score.
• The 4.0/4.1 generation and the 4.5+ generation are not drop-in. The 4.5+ reset changed pricing AND scores; treat pre-4.5 numbers as an older line that happens to share the brand prefix.
• Mythos data is Anthropic's announcement set. Independent reproductions had not landed in our index at last verification. Pricing and context window are unset (preview, no public SKU).
• Opus 4.6-thinking has 1M context; Opus 4.0/4.1/4.5/4.7 do not. This is unintuitive (newer is not strictly larger context) and is worth checking against the variant table before committing.
• Pricing on this page is the published API list price. Volume agreements and the various Anthropic-direct vs Bedrock vs Vertex paths can change the unit economics; list price is a calibration anchor, not the cost ceiling.

If you are making a procurement decision, the variant table on this page is the load-bearing artifact. Cross-check pricing against Anthropic's own docs and your cloud-provider's Bedrock/Vertex pricing before you commit.

When to reach for which alternative

• Open-weights deployment is a requirement: Claude is API-only. The conversation moves to open-weights families (Qwen3, DeepSeek, Llama). Compare on the specific benchmark that matches your workload; the cross-family comparison views in our index are designed for this.
• You need the cheapest competent reasoning at API scale: DeepSeek V4 Pro Thinking lands at QS 98.0 with $0.435 / $0.87 pricing in our index, which is roughly an order of magnitude cheaper than Sonnet 4.6 thinking at comparable quality-score position. Run a side-by-side on your eval before committing to either.
• Previous-generation Claude is already in production: the Sonnet 3.5 / Haiku 3.5 line is on the sibling claude-3-5 surface in our index. Anthropic still serves them, and for some chat-default workloads the migration cost to 4 may not be earned by the score delta. Run the comparison before assuming the upgrade.

Sources worth reading

• Anthropic API pricing: authoritative price list per model and tier
• Claude model docs: variant identifiers, context windows, modality coverage
• Anthropic news + announcements: release notes for new generations and pricing changes