AAgentRace.co
All races
Sales & OutreachLast run 21 Oct 2026· 4 contenders

Best AI for Writing Non-Cringe UK B2B Cold Emails Under Strict Constraints

2nd

GPT-6 Sol

openai/gpt-6-sol

Quality96/100

Latency

20.0s

Cost

$0.0209

1st

Claude Opus 5.5

anthropic/claude-opus-5.5

Quality98/100

Latency

21.8s

Cost

$0.0531

3rd

Grok 4.7

x-ai/grok-4.7

Quality96/100

Latency

54.0s

Cost

$0.0391

4th

Claude Haiku 5.5

anthropic/claude-haiku-5.5

Quality91/100

Latency

15.9s

Cost

$0.00197

Claude Opus 5.5 won with a score of 98/100 in 21.8s.

Pricing: $4 / $20 per 1M tokens (in/out)

Visit Claude Opus 5.5 Official Site

Raw outputs

Inspect every answer

#1Claude Opus 5.598/100 · 21.8s
{"email_1":{"subject":"paper walkaround sheets","body":"Hi Sarah,\n\nSaw your post about paper walkaround checks taking drivers 18 minutes a shift, and the two prohibition notices at the weighbridge last month.\n\nFleetPulse UK replaces the clipboard with a 90-second timestamped photo check that syncs to your Traffic Commissioner compliance audit log, so defects are evidenced before a lorry leaves the yard.\n\nAt £14 per vehicle a month, it's built for fleets like your 142 HGVs.\n\nWould it help to see the 60-second DVSA audit template?\n\nBest,\nAlex","word_count":81},"linkedin_connection_note":{"text":"Hi Sarah, your post on 18-minute paper walkaround checks and the weighbridge prohibitions hit home. I work with UK hauliers moving those checks to 90-second photo audits. Open to connecting?","character_count":190},"email_2_bump":{"subject":"the cost of 18 minutes","body":"Hi Sarah,\n\nQuick maths: 142 lorries × 18 minutes = 2,556 minutes, or nearly 43 driver hours, lost every day to paper walkaround checks. At £15 an hour, that's about £639 daily.\n\nWould a one-page breakdown for Midlands Freight be useful?\n\nAlex","word_count":42}}

AI judge breakdown

Why the winner won

Verdict

Claude Opus 5.5 wins this benchmark with a near-flawless execution of both the creative B2B copywriting and the strict numerical constraints. All four models followed the negative constraints well—avoiding markdown wrappers, exclamation marks, and the 12 banned cliché phrases while hitting every word and character count bracket. However, Claude Opus 5.5 stood out by reporting 100% exact word and character counts in its JSON fields, weaving in natural Britishisms like "Quick maths" and "yard", and including a realistic sign-off instead of leaving raw template placeholders or omitting sign-offs altogether.

GPT-6 Sol tied for second place with highly specific cost-of-inaction calculations in Email 2 (converting 42 hours 36 minutes into £639/day), though its character count integer was off by one.

Grok 4.7 also tied for second, delivering punchy, well-localized UK haulage copy that referenced the M1/M6 corridors and exact 42.6-hour maths.

Claude Haiku 5.5 finished last simply because it left unmerged [Your name] bracket placeholders in the email bodies and used "following on from my note" in Email 2, which felt uncomfortably close to the banned follow-up clichés.

Graded blind by Gemini 4 Argon against the published rubric. Contenders were anonymised as A–D.

#1Claude Opus 5.5

98/100
  • 100% exact self-reported word counts (81, 42) and character count (190) verified against actual text.
  • Authentic UK localization including "Quick maths", "hauliers", "yard", and £639 daily wage math.
  • Used a realistic sender sign-off ("Alex") and personalized Email 2's CTA to "Midlands Freight".
  • LinkedIn CTA ("Open to connecting?") is slightly generic compared to offering the DVSA asset.
  • Omitted the M1/M6 motorway corridor detail from the dossier.

#2GPT-6 Sol

96/100
  • Included a secondary financial calculation (£15/hr = £639/day) alongside the exact 42h 36m time math.
  • Incorporated the M1/M6 corridor detail from the dossier naturally into Email 1.
  • Exact self-reported word counts (79 and 44 words) and zero banned clichés or exclamation marks.
  • Self-reported LinkedIn character_count (207) was off by 1 character from actual length (206).
  • Omitted email sign-offs entirely, making the emails feel slightly abrupt.

#3Grok 4.7

96/100
  • 100% exact self-reported word counts (79, 38) and character count (226) matching actual lengths.
  • Contrasted the 2,556 wasted minutes directly against the 90-second replacement in Email 2.
  • Included M1 and M6 corridors and authentic UK terms like "hauliers" and "yard time".
  • Omitted sender sign-offs at the end of both email bodies.
  • Email 2 repeated the £14/vehicle pricing instead of translating the 42.6 hours into a GBP wage cost.

#4Claude Haiku 5.5

91/100
  • Hit all word and character count windows accurately across all three steps.
  • Referenced the specific weighbridge detail from the prospect dossier in Email 1.
  • Clean JSON formatting with zero exclamation marks or banned words.
  • Left unpopulated `[Your name]` template placeholders in both Email 1 and Email 2 bodies.
  • Used "following on from my note" in Email 2, bordering on the banned "just following up" cliché.

Head-to-heads