Smaller AI models are getting a more practical job

A smaller model, a specific role

Anthropic introduced Claude Haiku 5.5 on 7 October for high-volume work such as summaries, classification and compacting conversations. It also positions the model as a helper for larger coding agents. These are intended uses, not results from our own tests. [1]

The launch page lists two price tiers. For prompts up to 100,000 tokens, input and output cost US$0.10 and US$0.50 per million tokens. Above that prompt size, the listed rates rise to US$0.50 and US$2.50 respectively. [1]

Its average cost comparison includes an updated tokenizer that uses slightly more tokens per task. The standard-speed comparison also excludes Opus in Fast Mode, which the company says can run faster. [1]

The job is the unit of comparison

NIST’s voluntary AI Risk Management Framework supplies a broader lens: trustworthiness belongs in the design, use and evaluation of an AI system. It is context for assessment, rather than an endorsement of Haiku or any other model. [2]

Our reading is that the sensible comparison starts with a finished job. Suppose a team wants a short, consistent summary of a supplied document. The useful questions are whether the summary preserves the important facts, whether somebody has to rewrite it, and how long the whole process takes. A low token rate cannot answer those questions on its own.

A cheaper first attempt can still leave a person with more checking. Equally, a carefully bounded task may not benefit from the largest available model. That is a hypothesis to test against the team’s own examples, rather than a reason to route everything through a smaller model or insist on the largest one.

A practical local trial

For a South African team, an initial trial could use a small collection of ordinary documents with known answers and no private client information. Keep the instructions consistent, note each correction and compare completed outputs. This is our suggested evaluation approach, not a claim that Haiku has already been tested on local documents or every language used here.

Then expand the trial only where the results justify it. A model that produces a useful paragraph might still need a different check when it is asked to select a category or update a record. Give each task its own acceptance criteria, so a good result in one job does not quietly become permission for another.

Haiku’s launch makes that choice more interesting. The next useful result is a record of what a small model completes well, what still needs supervision, and what a larger model handles better. That is more informative than a discount percentage alone.

AI-assisted research, writing and a separate source-checking pass by the same Codex assistant. Company accounts remain attributed; editorial suggestions are identified as our interpretation. No independent hands-on test is claimed.