▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=38_6C0dkKmU
• Original title :: finally a good small model

Theo reviews Haiku 5.5, the first Anthropic small model he thinks is worth using. It costs about 75% less than Haiku 4.5 and is very cheap below a 100k-token context; above that the price goes up 5x. It also scores more than double Haiku 4.5 on most benchmarks. On cost per intelligence it is now close to GPT-6 Luna, though Luna and 6.1 Soul are still slightly better per dollar. Its value is as a subagent and an in-app tool for search, categorization and verification, not as the model you code with. In his own tests, Haiku driving a task alone boxed itself in. Opus orchestrating Haiku subagents worked well. The same release also halves Sonnet 5.5's cache-read price and gives Max subscribers monthly API credit.

Pricing and the 100k catch

• Haiku 5.5: $0.10/M input, $0.50/M output, $0.01/M cache reads, $0.125/M cache writes. That is roughly 20x cheaper than Sonnet 5.5 on input/output and 10x cheaper on cache reads.
• Above a 100k-token context every price goes up 5x (e.g. $2.50/M output). That is still half the old Haiku 4.5 price.
• Theo's reading: the cheap tier targets quick classification/summarization calls (competing with things like Jev). The surcharge stops Anthropic from going broke when Haiku runs long agent sessions in Claude Code.
• Workaround from Anthropic staff: set Haiku's auto-compact window to 100k in Claude Code. The setting is per model, so subagents stay in the cheap tier.
• In his Fish Slop game build, 47 of 51 requests went over 100k. The ~$1 run was mostly 9.7M cache-read tokens.
• Sonnet 5.5's cache reads dropped from $0.20 to $0.10/M, the same as 6.1 Soul. They used to cost as much as Opus's, which made Sonnet a poor deal for agentic work. Cache reads are 30–50% of an agentic bill.

Benchmarks

• More than double Haiku 4.5 on most benchmarks. Terminal Bench went from ~0 to almost 40%, versus 16.4% for GPT-6 Luna.
• Strong at computer use, well ahead of Luna. Speed is 100–200 TPS; Theo saw ~180 TPS in Claude Code.
• ThePrimeagen swapped it in with reasoning off and got worse pass rates than Luna. Anthropic models are bad without reasoning. If you want no reasoning, use Jev, or Luna if you need vision.
• Cost vs intelligence (Artificial Analysis data): Haiku 5.5 sits far left, near Luna. Opus 5.5 is about 5x the cost for a similar aggregate score.
• Best value per dollar across vendors: GPT-6 Luna → 6.1 Soul → Opus 5.5. Mimo V26 Pro is an underrated open-weight option. Gemini 4, Grok 4.7 and Gemini Flash add nothing.
• Within Anthropic, Opus + Haiku covers almost everything; Sonnet adds little even after the price cut. Never use Sonnet 5.5 on max reasoning ($7.60/task).
• Opus 5.5 on low costs only ~2x Haiku 5.5 on max per task.

Other launch details

• Max 20x subscribers get $200/month in API credit and 5x subscribers $100. You can use it to ship apps with Claude built in. Theo says Anthropic changed its plan after feedback.
• The Python and TypeScript SDKs now support computer use and browser use.

Where Haiku fits

• Anthropic's demo: Opus 5.5 alone did an egg-drop design task in 3.5 min, 25 attempts, $0.47. Opus plus 10 Haiku subagents did it in under 1 min, 86 attempts, $0.14.
• It shines where many cheap attempts plus verification beat one careful attempt: finding a file you can't grep for, exploring codebases, reading docs, checking theories, categorizing data.
• Writing code is not what makes agents expensive. The cost is in gathering context before and verifying after, plus retries. Once the context is cached, emitting the code file costs pennies even on Fable. So a dumb model for the coding step saves little.
• A model that fails can cost more: if Opus solves something in 100k tokens and Haiku burns 1M before giving up, Haiku was the more expensive choice.
• Theo is removing the "never use Haiku" rules from his CLAUDE.md and letting Opus decide when to use Haiku subagents.

Demos

• A community video made programmatically for $0.60 in under 20 minutes. Front-end designs are decent and far better than Grok 4.7, but only with Anthropic's design skill; without it the output is generic LLM design.
• Fish Slop game: roughly Grok-tier. It had no mouse look, broken corners and a wrong attack timing. Its lows are not as low as Astra's, but its highs are a tenth as high.
• PR triage on ~1,500 T3 Code PRs, Haiku alone: it fired parallel GitHub fetches and exhausted the 5,000/hr rate limit. It saved error bodies as data, then said it would wait without scheduling a wakeup, and stalled.
• Same task with Opus orchestrating ~100 Haiku subagents (15 PRs each): it planned around the scale and the rate limits.

Verdict

• Haiku 5.5 is "sometimes worth using", which is new for an Anthropic small model. Use it inside apps and as a subagent; don't pick it as your coding model.
• Theo's setup: Opus writes the code, 6.1 Soul audits it before the PR (a different model family brings a different perspective), and Haiku does bulk reading and categorization.
• He thinks Anthropic's 5.5 lineup leads the field and that OpenAI should be worried. He hopes for a Fable 5.5.