Claude 4.5 Sonnet Is Here: The Coding Model That Actually Listens

Anthropic released Claude 4.5 Sonnet a few days ago with surprisingly little fanfare, which is on brand. No splashy launch event, just a model card update and a price drop. After running it through my normal battery of tasks, I think this is the most underrated release of the year so far. It will not headline benchmarks the way Gemini 3.1 Pro did, but for everyday coding and writing work it has quietly become my default.

Here is what changed and what it means if you live inside an editor or a terminal all day.

The headline upgrades

Coding is the obvious focus. On a private set of refactor tasks I keep around, Claude 4.5 Sonnet completed eighteen out of twenty without intervention. Claude 4 Sonnet did fourteen. The difference is mostly in following instructions exactly. When you say do not touch the test files, it actually does not touch the test files.

Tool calling is more reliable. Earlier Claude versions would sometimes invent a function signature or call a tool with the wrong argument shape. The new model reads the tool schema carefully and almost never improvises. For anyone building agents this is a real quality of life improvement.

Context handling improved on the long end. The window is still two hundred thousand tokens, but the model now actually uses information from the early parts of a long conversation instead of forgetting it after the halfway mark.

The voice is still distinctly Claude. Calm, direct, slightly formal. If you found Claude 4 a bit too eager to add caveats, you will be happy to know the caveats are dialed back here.

What it does best

Pair programming inside a real codebase. Drop it into Cursor or Zed with the full repository in context and it will navigate, read, and edit with a level of restraint that most other models lack. It will not rewrite half your file when you asked for a one line change.

Long form writing where tone matters. Technical documentation, internal memos, considered email replies. The output reads like it was written by a careful human, not a machine trying to sound human.

Reading and summarizing dense material. Legal text, research papers, product specs. It catches nuance and flags ambiguity instead of papering over it.

Where it still trips

Math and numerical reasoning at the harder end. It is fine on day to day arithmetic and stats, but on competition style problems Gemini 3.1 Pro and GPT-5.2 both pull ahead.

Multimodal reasoning. Image understanding works but feels a generation behind Gemini. If your workflow leans heavily on screenshots or charts, this is not the model for you.

Creative fiction with a strong voice. Output is competent but rarely surprising. For marketing copy with personality, you will get more interesting results from a tuned GPT-5.2 prompt.

Pricing matters here

Anthropic dropped the input price by roughly thirty percent at the same time. For high volume coding work this is significant. If you were burning through budget on Claude 4 Sonnet, the same workload now costs less and produces better output. That combination is rare.

The Sonnet tier remains the sweet spot for most production use. Opus is overkill for coding tasks where Sonnet does the job. Haiku is still the right pick for high volume classification and routing.

My take

The frontier model conversation right now is dominated by Gemini 3.1 Pro and GPT-5.2 trading punches on benchmarks. Claude 4.5 Sonnet is not in that fight and does not pretend to be. What it is doing instead is being the model you can trust to do exactly what you asked, in code or in prose, without drama.

For solo operators and small teams shipping real products, that reliability is worth more than a few extra benchmark points. I have moved my default coding agent to Claude 4.5 Sonnet and kept Gemini for research and GPT-5.2 for creative work. That stack costs less than what I was running last month and produces better output. Hard to argue with.

If you have not tried it yet, spend an afternoon with it inside your editor. You will probably notice the difference within an hour.