Compare / Claude Haiku 5.5 vs Mercury 3

Claude Haiku 5.5 vs Mercury 3

Which is cheaper, which is more capable, and which does better at each kind of work, from 51 results on 4 sources that measured both models.

AnthropicClaude Haiku 5.5 Released 7 Oct 2026 · $0.10 in · $0.50 out
InceptionMercury 3 Released 19 Oct 2026 · Not on sale
Claude Haiku 5.5 wins 2814 too close to callMercury 3 wins 9

Of the 51 results both models have, the better value on each. Results whose reported ranges overlap are too close to call.

Results as of 19 November 2026 (catalogue release 2026-11-19-3608133e8331).

Claude Haiku 5.5 or Mercury 3? The short answer

Pick Claude Haiku 5.5: it scores higher (Spring Prompt overall 38 against 33) and costs less.

  • Spring Prompt overall: Claude Haiku 5.5 38, Mercury 3 33 (out of 100; how the score works).
  • Claude Haiku 5.5 is cheaper to run: $0.009 against $0.012 per deck on DeckBench (22% less).
  • Mercury 3 is newer: released 19 Oct 2026, against 7 Oct 2026 for Claude Haiku 5.5.
  • On our own benchmarks, on the results both models have: Claude Haiku 5.5 does better for decks from an analysis and user surveys; Mercury 3 for marketing planning.

Head to head

38
Spring Prompt overall
33

Every fact side by side

Claude Haiku 5.5Mercury 3
Price per million tokens$0.10 in · $0.50 outNot on sale
Blended (3 in : 1 out)$0.20–
Cheapest host, blended$0.20 (Amazon Bedrock)–
Cost per deck (DeckBench)$0.009$0.012
Intelligence Index43.4–
Spring Prompt overall3833
Output speed, tokens a second240–
Context window1,000,000 tokens–
Released7 Oct 202619 Oct 2026 (newer)
WeightsClosedClosed
Developer based inthe USthe US

Prices are the developer's list price where we have read it, otherwise the typical host on OpenRouter; see every model's price. Intelligence and speed from Artificial Analysis.

Which is better for what

Each model's rank among the models on sale today, for each kind of work, from our own benchmarks and licensed sources. Further right is better; the better rank is in green.

Product listingsTurning a sparse product feed and photos into listings that can go live 23 of 35 not measured
Decks from an analysisTurning a finished analysis into a deck you could present as it is 19 of 35 30 of 36
User surveysPlanning a user survey and reading its results without being misled 9 of 26 21 of 27
Marketing planningPlanning a year of ad spend without overspending 33 of 35 30 of 36
Agents and tool useMulti-step tasks with tools: support desks, coding agents, function calls 28 of 159 not measured
Reasoning and knowledgeHard questions across science, maths and general knowledge 25 of 187 not measured
Speed under pressureGood decisions against a real clock (fast chess) 24 of 38 16 of 39

Their slides, side by side

The same brief for both: Manchester store investment review for Hearthside Coffee, an invented company. First slide of each deck from DeckBench.

First slide of Claude Haiku 5.5's deck for Hearthside Coffee
Claude Haiku 5.5 · deck rating 1,120 · needs work
First slide of Mercury 3's deck for Hearthside Coffee
Mercury 3 · deck rating 726 · needs work

Every slide, side by side →

From our research

Every shared result

Each model's best result on every measure both have, grouped by benchmark, with the setting that produced it. Under each measure, every model that benchmark has measured is a grey tick, best to the right, so you can see where the two sit in the field. The headline measures of each benchmark come first; the rest are under "Show all". Green marks the better value; where a source reports ranges that overlap, neither is marked.

MeasureClaude Haiku 5.5Mercury 3
BulletBench measured by Spring Prompt · 3 results
Cost per game · Bullet 60sUS dollars, lower is better 23 · 24 of 62 measured $0.0019no reasoning $0.0020no reasoning
Ladder Elo · Bullet 60sladder Elo, higher is better 50 · 33 of 62 measured 310no reasoning 491no reasoning
Games lost on time · Bullet 60s% of games, lower is better 36 · 1 of 62 measured 37.5%no reasoning 0.0%no reasoning
DeckBench measured by Spring Prompt · 3 results
Cost per deckUS dollars, lower is better 4 · 6 of 36 measured $0.0091 $0.0116
Deck ratingrating, higher is betterThe rating weighs each win by how strong the other deck was; the win rate counts wins against whichever decks a model happened to meet, so the two can disagree. 19 · 30 of 36 measured 1,120 726
Accurate decks% of tasks, higher is better 22 · 31 of 36 measured 33.3% 0.0%
ROASBench measured by Spring Prompt · 3 results
Overall scorescore out of 100, higher is better 34 · 30 of 36 measured 24.1 26.1
Audience scorescore out of 100, higher is better 30 · 33 of 36 measured 48.4 43.2
Business scorescore out of 100, higher is better 32 · 30 of 36 measured 13.5 17.6
SurveyBench measured by Spring Prompt · 3 results
Analysis ratingrating, higher is betterThe rating weighs each win by how strong the other analysis was; the win rate counts wins against whichever analyses a model happened to meet, so the two can disagree. 9 · 21 of 27 measured 1,233 735
Cost per taskUS dollars, lower is better 4 · 7 of 27 measured $0.0185 $0.0216
Analysis scorepoints of 100, higher is better 11 · 24 of 27 measured 88.4 42.2
Show all 51 results (39 more)
MeasureClaude Haiku 5.5Mercury 3
BulletBench measured by Spring Prompt · 2 results
Invalid moves · Bullet 60s% of moves, lower is better 1 · 1 of 62 measured 0.0%no reasoning 0.0%low reasoning
Median move time · Bullet 60smilliseconds, lower is better 43 · 3 of 62 measured 1.7 sno reasoning 0.4 sno reasoning
DeckBench measured by Spring Prompt · 14 results
Caveat dropped% of tasks, lower is better 16 · 16 of 36 measured 16.7% 16.7%
Clean layout% of tasks, higher is better 10 · 10 of 36 measured 66.7% 66.7%
Design quality% of the maximum, higher is better 22 · 25 of 36 measured 61.1% 58.5%
Draft figure quoted% of tasks, lower is better 21 · 21 of 36 measured 33.3% 33.3%
Findings missing% of tasks, lower is better 25 · 25 of 36 measured 16.7% 16.7%
Head-to-head win rate% of comparisons, higher is betterThe rating weighs each win by how strong the other deck was; the win rate counts wins against whichever decks a model happened to meet, so the two can disagree. 15 · 30 of 36 measured 62.1% 35.2%
Layout defects% of tasks, lower is better 10 · 10 of 36 measured 33.3% 33.3%
Misleading metric used% of tasks, lower is better 31 · 31 of 36 measured 16.7% 16.7%
Presentable decks% of tasks, higher is better 9 · 17 of 36 measured 33.3% 0.0%
Recommendation late or wrong% of tasks, lower is better 22 · 33 of 36 measured 33.3% 83.3%
Slides needing work% of tasks, lower is better 5 · 12 of 36 measured 16.7% 33.3%
Time per deckseconds, lower is better 7 · 12 of 36 measured 62.1 83.3
Unsupported claims% of tasks, lower is better 26 · 17 of 36 measured 33.3% 16.7%
Unsupported numbers% of tasks, lower is better 28 · 28 of 36 measured 16.7% 16.7%
ROASBench measured by Spring Prompt · 6 results
Consistency scorescore out of 100, higher is better 33 · 31 of 36 measured 10.6 15.1
Contribution profitsimulated US dollars, higher is better 32 · 30 of 36 measured $32,146 $104,868
Cost of a runUS dollars, lower is better 6 · 8 of 36 measured $0.0313 $0.0533
Months over budgetmonths of 12, lower is better 1 · 1 of 36 measured 0 0
Planning scorescore out of 100, higher is better 21 · 35 of 36 measured 54.8 53.8
Return on ad spendprofit per $1 spent, higher is better 32 · 30 of 36 measured $0.03 $0.10
SurveyBench measured by Spring Prompt · 17 results
Head-to-head win rate% of comparisons, higher is betterThe rating weighs each win by how strong the other analysis was; the win rate counts wins against whichever analyses a model happened to meet, so the two can disagree. 8 · 18 of 27 measured 59.9% 42.4%
Leading or double-barrelled questions% of tasks, lower is better 12 · 21 of 27 measured 33.3% 55.6%
Mistaken next steps% of tasks, lower is better 13 · 13 of 27 measured 22.2% 22.2%
Numbers right% of answers, higher is better 1 · 27 of 27 measured 100.0% 63.3%
Objectives unanswerable% of tasks, lower is better 1 · 1 of 27 measured 0.0% 0.0%
Precision ignored% of tasks, lower is better 1 · 1 of 27 measured 0.0% 0.0%
Research scorepoints of 100, higher is better 10 · 23 of 27 measured 90.2 56.8
Sound research% of tasks, higher is better 13 · 17 of 27 measured 11.1% 0.0%
Stakeholder's view presumed% of tasks, lower is better 1 · 18 of 27 measured 0.0% 22.2%
Survey plan scorepoints of 100, higher is better 8 · 18 of 27 measured 94.5 90.8
Survey structure% of tasks, lower is better 12 · 12 of 27 measured 22.2% 22.2%
Time per taskseconds, lower is better 13 · 16 of 27 measured 147 194
Traps handled% of traps, higher is better 11 · 25 of 27 measured 94.2% 60.7%
Traps missed% of tasks, lower is better 4 · 19 of 27 measured 22.2% 88.9%
Unsupported findings% of tasks, lower is better 16 · 11 of 27 measured 77.8% 44.4%
Wrong numbers% of tasks, lower is better 1 · 21 of 27 measured 0.0% 44.4%
Wrong recommendation% of tasks, lower is better 5 · 24 of 27 measured 11.1% 77.8%

Quick answers

Is Claude Haiku 5.5 better than Mercury 3?

Pick Claude Haiku 5.5: it scores higher (Spring Prompt overall 38 against 33) and costs less.

Which is cheaper, Claude Haiku 5.5 or Mercury 3?

Claude Haiku 5.5 is cheaper to run: $0.009 against $0.012 per deck on DeckBench (22% less).

Which is better for decks from an analysis, Claude Haiku 5.5 or Mercury 3?

Claude Haiku 5.5 ranks 19 of 35 models on sale for turning a finished analysis into a deck you could present as it is, against 30 for Mercury 3.

Which is better for user surveys, Claude Haiku 5.5 or Mercury 3?

Claude Haiku 5.5 ranks 9 of 26 models on sale for planning a user survey and reading its results without being misled, against 21 for Mercury 3.

Which is better for marketing planning, Claude Haiku 5.5 or Mercury 3?

Mercury 3 ranks 30 of 35 models on sale for planning a year of ad spend without overspending, against 33 for Claude Haiku 5.5.

Which is better for speed under pressure, Claude Haiku 5.5 or Mercury 3?

Mercury 3 ranks 16 of 38 models on sale for good decisions against a real clock (fast chess), against 24 for Claude Haiku 5.5.

Run Claude Haiku 5.5 and Mercury 3 on your own prompt.

Benchmarks aren't your task. Test both side by side on your prompt, with every answer checked against your rules and the cost of each one.