Claude 3.5 Sonnet

Claude 3.5 Sonnet

Quick answer: Claude 3.5 Sonnet (October 2024, claude-3-5-sonnet-20241022) is Anthropic's mid-tier model that set the standard for cost-efficient frontier coding and reasoning at its release. It scores 49.0% on SWE-bench Verified and 65.0% on GPQA Diamond, operates on a 200K-token context window, and is priced at $3/$15 per million tokens — delivering Opus-class performance at Sonnet speed.

At a Glance

Where Claude 3.5 Sonnet leads

• Strong coding benchmark — 49.0% on SWE-bench Verified at Sonnet pricing when it was released
• GPQA Diamond 65.0% — graduate-level scientific reasoning at mid-tier cost
• 200K-token context with solid long-document fidelity for contracts, codebases, and research

Where it lags

• Superseded by Claude Sonnet 4 and later Sonnet 4.6 for new builds; use 3.5 only if you need the specific older checkpoint
• 8,192-token output limit is lower than newer models (up to 128K on Sonnet 4.6)
• No extended thinking or adaptive effort modes — single-pass inference only

Best for: teams maintaining existing Claude 3.5 Sonnet integrations, or workflows that need a well-characterised, stable checkpoint for regression testing and fine-tuning baselines.

What Claude 3.5 Sonnet Is

Claude 3.5 Sonnet was launched by Anthropic on June 20, 2024, and updated to its current checkpoint (claude-3-5-sonnet-20241022) on October 22, 2024. At release, it was the most capable Anthropic model available, outperforming Claude 3 Opus across nearly every benchmark while running at twice the speed and a fraction of the cost. The October 2024 update added computer use capability — the ability to interact with desktop applications and browsers via screenshots — making it the first frontier model to ship that feature in general availability.

The model sits in Anthropic's "Sonnet" tier: balanced intelligence for high-volume production use. It is a dense transformer trained on text and images, closed-source, and available only via API. Knowledge cuts off at April 2024.

From a Benchgen perspective, Claude 3.5 Sonnet is a historically important checkpoint and a common baseline for agent quality comparisons. Many teams that started building agents in 2024 used it as their primary backbone, which means Benchgen trajectory data collected during that period will predominantly reflect its behavior. For new builds, Sonnet 4.6 is the current recommendation — but understanding 3.5 Sonnet's behavioral profile remains useful when interpreting legacy evaluation results or comparing across generations.

Specifications

Pricing

Prompt caching reduces input costs by up to 90%; batch processing reduces both by 50%. Source: Anthropic pricing page.

Context Window

Claude 3.5 Sonnet has a 200,000-token context window — roughly 300 pages of text in a single request — with a maximum output of 8,192 tokens. That is enough to hold a large codebase, a lengthy contract stack, or a full research paper at once. The October 2024 version added computer use, letting the model read the screen and take mouse/keyboard actions inside the context of a task.

Public Benchmark Scores

Scores are reported by Anthropic and shown for context. They are not Benchgen measurements. See the source for harness and methodology details.

Claude 3.5 Sonnet vs Alternatives

At the same $3/$15 price point, Sonnet 4.6 substantially outperforms 3.5 Sonnet on every current benchmark. The only reasons to stay on 3.5 Sonnet are version-pinning requirements, existing fine-tunes, or the need to compare against a fixed historical baseline. (Rival scores sourced from respective provider announcements; not Benchgen measurements.)

How Claude 3.5 Sonnet Performs on Real Agent Tasks

Claude 3.5 Sonnet's 49.0% SWE-bench Verified score was a high-water mark when it was published in mid-2024 — a strong signal that the model could handle realistic, multi-file software engineering tasks rather than just toy coding puzzles. The computer use feature added in October 2024 extended this to browser and desktop automation, and Anthropic's own internal agentic coding evaluation reported 64% task completion.

What those numbers reflect is consistent mid-tier agentic reliability: the model makes reasonable tool calls, recovers from errors with moderate reliability, and handles instruction chains of 5–15 steps well. Where it tends to break down is in very long-horizon tasks — 30+ step workflows with many state-dependent decisions — where the absence of extended thinking and the 8,192-token output cap can cause the model to truncate or repeat steps. For teams evaluating this with Benchgen, expect strong Tool-Call Accuracy and Goal Completion scores on sub-15-step tasks, with regression risk on longer workflows.

Use Claude 3.5 Sonnet via API

from anthropic import Anthropic

client = Anthropic(api_key="YOUR_API_KEY")

message = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Review this Python function for bugs..."}],
)
print(message.content[0].text)

Related Models & Benchmarks

• Compare: Claude Sonnet 4.6 · Claude Opus 4.6 · GPT-4o · Gemini 1.5 Pro
• Benchmarks: SWE-bench Verified · MMLU-Pro

Frequently Asked Questions

Specs and scores sourced from Anthropic's Claude 3.5 Sonnet announcement and the model card addendum. Third-party benchmark scores attributed inline. Last updated 2026-06-19.