WhatsApp PersonalBusiness
Recruitly LogoRecruitly

Anthropic: Claude Haiku 5.5

Text · via Direct API · 18 Nov 2026

Claude Haiku 5.5 is a large language model from Anthropic, released in October 2026. It is built for high-volume, latency-sensitive work such as classification, routing, extraction, and subagent tasks. It supports adaptive thinking with an effort parameter, a 1M token context window, and up to 128k output tokens.

Anthropic's page for Claude Haiku 5.5 ↗

Score
44
out of 100
Jobs passed
15/19
4 failed
Cost per task
$0.0012
measured
Speed
8.5s
median
Price per million
$0.1 / $0.5
1,000,000 context

Results by job

Cover note — strong candidate who needs sponsorship the client cannot give
sponsorship (and the £92k ask) flagged to the recruiter only; the client email never mentions visa, sponsorship or her pay; "Hi Harriet," on its own line; no sign-off
0%09.7s$0.0011
Cover note — three candidates, one a mismatch; two recipients
Rebecca flagged unsuitable and not pitched; Elena and Kwame suitable; "Hi Graham and Andrew," (not Sarah); the £45k vs £95-105k clash in objections, not in the email
0%08.9s$0.0012
CV parse — three messy pages, roles after education, a referee's contacts
5 jobs newest first in yyyy-MM (Sept '21 → 2021-09, current end ""), no career-break row, her own email and phone digits, not the referee's, no LinkedIn URL
100%446.7s$0.0011
Data enrich — parse + cleanup of a messy record with a CV
names cased, mobile E.164, placeholders gone, 03/04/2023 kept, years from dates (~11) not the CV's 15, London not Manchester, no gender guessed
0%012.9s$0.0019
Email intent — CV-share reply: one yes, one pass, a resend, an unlisted name, a reschedule, a leave date
GENUINE_REPLY; Priya INTERESTED, Tom NOT_INTERESTED, Aleksander unjudged, Grace not rejected, no Marek; asks interview + info + more + reschedule; proposedTimes 19th 14:00, 20th 10:00, 23rd 14:00 in that order; no 18th/24th/1 Dec
100%725.1s$0.0009
LinkedIn search cards → JSON array, a full batch of ten noisy cards
ten objects in card order; doubled name, pronouns and CIPD dropped; company from Current: (Brightwave Group, Meridian Search, Kite Health), null where only a Past: employer; masked surnames kept as K. and H
100%6210.2s$0.0017
KB answer — team signature among a stale twin and look-alikes
answers from the company template article (docs 2 and 4): Settings > Company > Email Branding, Apply to all users, both its screenshots; the archived doc 1 is not given as current
100%952.9s$0.0004
KB answer — nothing in the passages answers
[[KB_NO_ANSWER]]: no passage mentions Xero; the Zoho Books article (and the earlier Zoho answer) does not answer a Xero question
100%61.1s$0.0002
Meeting notes — client briefing called "Catch up", corrections mid-sentence
client_briefing; two roles at 80-88 base + 10% (not 90, not three); 4 Dec (not the 3rd); shortlist by Tuesday (not Wednesday); scorecard = Priya, spec = Marcus; ring Daniel = call, shortlist email = not; no reference-check action
100%847.7s$0.0011
Match rationale — same title, wrong industry
all 20 jobs, indexes as labelled, the utility roles (2, 3) above the same-title scaffolding, SaaS, brewery and events roles, ordered best first, no made-up years or figures
0%010.5s$0.0014
Application review — strong match, full sheet
high scores, every scorecard id, all 3 roles in yyyy-MM, contact copied verbatim, mentoring left partial
100%6814.0s$0.0022
Application review — irrelevant CV
every score 0-25 (the prompt's rule for an irrelevant CV), Reject, technical lines all 'no'
100%426.0s$0.0011
Application review — job-hopper with a call note
Experience & Stability section quoting 12.4 months, experience in the 40-59 band, stability risk/moderate, no invented billings
100%5019.3s$0.0026
Application review — AI instruction overrides a strong CV
the AI instruction wins: overall ≤40, right-to-work line 'no', Reject, while skills stay high
100%4612.3s$0.0020
LinkedIn result analysis — job brief with deal-breakers and scorecard
every profile line judged (4 must, 3 preferred, 2 deal-breakers), both deal-breakers not_met, licence not 'met', no salary/notice pick, emails open 'Hi Aoife,' in the agency's voice
100%3023.9s$0.0033
Support compose — customer asks for what only the team-only notes hold
options tell her Mark finds Recruitly by name and grants admin consent (customer article), and send neither the client ID nor a consent link; first name Sarah-Jane only; no app registration / client ID / SupportOps / Ravi; no raw tool fields; no invented price or link; guide names the customer article or nothing
100%545.4s$0.0008
Support triage — thank-you while the engineers still owe a fix
not resolved (engineers promised, not delivered), no auto_resolve, needsHuman true, urgency high, intent bug, a usable topic
100%702.6s$0.0004
Writing assist — messy draft full of facts
typos fixed, £55k / 10% / 2 days / 3 months / 550 kept, paragraphs kept, nothing added
100%391.1s$0.0001
Writing assist — reply that ignores the candidate's question
spelling fixed, and the salary/remote question stays unanswered — nothing from the thread is borrowed
100%620.9s$0.0001