1gpt-5.4 highLeader
GPT-5.4 · high
Cost / call
$0.07
Range
87.0–96.0 · n=25
salesevals.com/Evaluated Apr 30, 2026
Which models know sales?
Eighteen model configurations coach the same 25 synthetic sales calls. Each call has hidden ground truth. A judge scores every coaching note from 0–100 on whether it found the real strengths, flaws, and next moves.
- Calls
- 25
- Models
- 18
- Evaluations
- 450
- Mean
- 89.8
25 calls · 450 evaluationsMetric: OverallBuild-time static dataEvals completed Apr 30, 2026
Ranked by overall
Leaderboard
Average 89.8 across 450 evaluations.
92.0
2gpt-5.5 none
GPT-5.5 · none
Cost / call
$0.17
Range
76.0–96.0 · n=25
3gpt-5.5 xhigh
GPT-5.5 · xhigh
Cost / call
$0.18
Range
76.0–97.0 · n=25
4gpt-5.4 xhigh
GPT-5.4 · xhigh
Cost / call
$0.07
Range
82.0–96.0 · n=25
5gpt-5.5 high
GPT-5.5 · high
Cost / call
$0.18
Range
72.0–97.0 · n=25
6gpt-5.5 medium
GPT-5.5 · medium
Cost / call
$0.17
Range
82.0–97.0 · n=25
7gpt-5.4 medium
GPT-5.4 · medium
Cost / call
$0.07
Range
81.0–96.0 · n=25
8gpt-5.4 none
GPT-5.4 · none
Cost / call
$0.07
Range
82.0–96.0 · n=25
9gpt-5.5 low
GPT-5.5 · low
Cost / call
$0.17
Range
84.0–96.0 · n=25
10gpt-5.4 low
GPT-5.4 · low
Cost / call
$0.07
Range
78.0–96.0 · n=25
11opus 4.7 max
Claude Opus 4.7 · max
Cost / call
$0.15
Range
72.0–96.0 · n=25
12opus 4.7 high
Claude Opus 4.7 · high
Cost / call
$0.12
Range
71.0–95.0 · n=25
13opus 4.7 xhigh
Claude Opus 4.7 · xhigh
Cost / call
$0.13
Range
75.0–95.0 · n=25
14opus 4.7 medium
Claude Opus 4.7 · medium
Cost / call
$0.10
Range
72.0–95.0 · n=25
15sonnet 4.6
Claude Sonnet 4.6 · default
Cost / call
$0.10
Range
84.0–95.0 · n=25
16opus 4.7 low
Claude Opus 4.7 · low
Cost / call
$0.09
Range
70.0–96.0 · n=25
17deepseek v4 pro
DeepSeek V4 Pro · default
Cost / call
$0.0047
Range
70.0–95.0 · n=25
18gemini 3.1 pro previewTrailing
Gemini 3.1 Pro Preview · default
Cost / call
$0.03
Range
67.0–93.0 · n=25
Filtered mean89.8
| # | Model | Avg | Cost / call | Distribution | Min | Max | n |
|---|---|---|---|---|---|---|---|
| 1 | gpt-5.4 highLeader GPT-5.4 · high | 92.0 | $0.07 | 87.0 | 96.0 | 25 | |
| 2 | gpt-5.5 none GPT-5.5 · none | 92.0 | $0.17 | 76.0 | 96.0 | 25 | |
| 3 | gpt-5.5 xhigh GPT-5.5 · xhigh | 92.0 | $0.18 | 76.0 | 97.0 | 25 | |
| 4 | gpt-5.4 xhigh GPT-5.4 · xhigh | 92.0 | $0.07 | 82.0 | 96.0 | 25 | |
| 5 | gpt-5.5 high GPT-5.5 · high | 91.7 | $0.18 | 72.0 | 97.0 | 25 | |
| 6 | gpt-5.5 medium GPT-5.5 · medium | 91.7 | $0.17 | 82.0 | 97.0 | 25 | |
| 7 | gpt-5.4 medium GPT-5.4 · medium | 90.9 | $0.07 | 81.0 | 96.0 | 25 | |
| 8 | gpt-5.4 none GPT-5.4 · none | 90.8 | $0.07 | 82.0 | 96.0 | 25 | |
| 9 | gpt-5.5 low GPT-5.5 · low | 90.8 | $0.17 | 84.0 | 96.0 | 25 | |
| 10 | gpt-5.4 low GPT-5.4 · low | 90.3 | $0.07 | 78.0 | 96.0 | 25 | |
| 11 | opus 4.7 max Claude Opus 4.7 · max | 90.2 | $0.15 | 72.0 | 96.0 | 25 | |
| 12 | opus 4.7 high Claude Opus 4.7 · high | 89.6 | $0.12 | 71.0 | 95.0 | 25 | |
| 13 | opus 4.7 xhigh Claude Opus 4.7 · xhigh | 89.4 | $0.13 | 75.0 | 95.0 | 25 | |
| 14 | opus 4.7 medium Claude Opus 4.7 · medium | 89.0 | $0.10 | 72.0 | 95.0 | 25 | |
| 15 | sonnet 4.6 Claude Sonnet 4.6 · default | 88.8 | $0.10 | 84.0 | 95.0 | 25 | |
| 16 | opus 4.7 low Claude Opus 4.7 · low | 87.6 | $0.09 | 70.0 | 96.0 | 25 | |
| 17 | deepseek v4 pro DeepSeek V4 Pro · default | 85.8 | $0.0047 | 70.0 | 95.0 | 25 | |
| 18 | gemini 3.1 pro previewTrailing Gemini 3.1 Pro Preview · default | 82.4 | $0.03 | 67.0 | 93.0 | 25 | |
| Filtered mean | 89.8 | ||||||
Top model
92.0avg
gpt-5.4 high
87.0–96.0 across 25 calls
Hardest call
79.7avg
MicrosoftCostco Wholesale
Costco Wholesale Proof-of-concept readout for analytics and productivity workflow with Microsoft
Product demomixed
Reasoning effort
+2.6pts
Claude Opus 4.7: low → max
87.6 → 90.2 as effort scales
How this benchmark was made
Methodology
The calls are synthetic, immutable benchmark cases with hidden coaching ground truth.
01
Generate cases
Each case starts with two companies, a call type, duration, quality target, web research, and role notes.
02
Write turn by turn
The transcript is produced through many structured LLM calls, one speaker turn at a time, using persona behavior and hidden coaching signals.
03
Judge semantically
Coach models see only the visible case. A judge model sees hidden ground truth and scores grounded coaching quality.
- Calls
- 25
- Model configs
- 18
- Judged runs
- 450
- Score axes
- 8
25 sales calls in the benchmark
Browse the calls
Pick a call to see the hidden answer key and how each model scored on it.
Discovery95.4
CollibraBerkshire Hathaway
Berkshire Hathaway Data governance discovery across decentralized business units with Collibra
flawed·33mEasiest
Competitive displacement94.3
StripePave
Pave Pricing and packaging objection call with Stripe
flawed·18m
Discovery94.1
VercelMercury
Mercury First discovery for frontend platform consolidation with Vercel
flawed·22m
Discovery94.0
AtlassianDelta Air Lines
Delta Air Lines Enterprise discovery for service management modernization with Atlassian
flawed·31m
Product demo93.7
MongoDBWayfair
Wayfair Integration deep dive for catalog modernization with MongoDB
excellent·58m
Renewal save93.7
TwilioThe Home Depot
The Home Depot Renewal save call after usage and support concerns with Twilio
flawed·42m
Product demo93.2
Palo Alto NetworksApple
Apple Technical security review for zero trust architecture with Palo Alto Networks
excellent·66m
QBR92.4
AmplitudeDuolingo
Duolingo Renewal QBR and expansion planning with Amplitude
excellent·52m
Discovery92.0
OpenAICVS Health
CVS Health AI contact-center transformation discovery with OpenAI
excellent·61m
Discovery91.8
GitHubRippling
Rippling Product-led expansion discovery for developer workflow with GitHub
excellent·41m
Discovery91.1
WorkdayMcKesson
McKesson HR transformation qualification and stakeholder mapping with Workday
flawed·27m
Product demo90.9
AnthropicExxonMobil
ExxonMobil AI governance and safety review for energy operations with Anthropic
mixed·39m
Product demo90.8
CrowdStrikeTarget
Target Security architecture review for endpoint consolidation with CrowdStrike
excellent·63m
Product demo90.4
DatadogLinear
Linear Technical demo for observability and incident response with Datadog
excellent·34m
Product demo90.4
ElasticJPMorgan Chase
JPMorgan Chase Technical workshop for search and observability consolidation with Elastic
excellent·74m
Discovery89.3
NVIDIAWalmart
Walmart Executive discovery for AI infrastructure and store operations with NVIDIA
excellent·57m
Discovery89.1
HashiCorpAmazon
Amazon Cloud operating model discussion for internal platform teams with HashiCorp
flawed·26m
Competitive displacement88.6
ServiceNowFord Motor Company
Ford Motor Company Procurement negotiation for workflow automation with ServiceNow
mixed·35m
Product demo87.0
SnowflakeToast
Toast Data platform proof-of-concept kickoff with Snowflake
flawed·44m
Competitive displacement85.8
CloudflareCanva
Canva Competitive displacement discovery for edge security with Cloudflare
flawed·47m
Product demo85.8
FigmaThe Walt Disney Company
The Walt Disney Company Design collaboration demo with brand and asset workflow discussion with Figma
mixed·49m
QBR85.2
OktaSweetgreen
Sweetgreen Executive alignment for identity modernization with Okta
mixed·38m
Renewal save84.9
SalesforceUnitedHealth Group
UnitedHealth Group Healthcare CRM expansion objection handling with Salesforce
mixed·46m
Product demo82.5
SnykRunway
Runway Security review before developer-tool rollout with Snyk
mixed·29m
Product demo79.7
MicrosoftCostco Wholesale
Costco Wholesale Proof-of-concept readout for analytics and productivity workflow with Microsoft
mixed·55mHardest
By model family
Does spending more on reasoning help?
For families with multiple reasoning settings, this shows the average score at each level.
Claude Opus 4.7
+2.6 pts
low
87.6
baseline
medium
89.0
+1.3 pts
high
89.6
+1.9 pts
xhigh
89.4
+1.8 pts
max
90.2
+2.6 pts
GPT-5.4
+1.1 pts
none
90.8
baseline
low
90.3
-0.5 pts
medium
90.9
+0.1 pts
high
92.0
+1.2 pts
xhigh
92.0
+1.1 pts
GPT-5.5
no change
none
92.0
baseline
low
90.8
-1.2 pts
medium
91.7
-0.3 pts
high
91.7
-0.3 pts
xhigh
92.0
+0.0 pts