Executive verdict
Grok 4.5 is a credible fast deep-work candidate, and vision is not a blocker.
It completed the three StrategyOS scenarios 6.7× faster than Doctrine, but cost
48.5% more in aggregate. Doctrine still produced the stronger consulting analysis
on the substantive Sprinta PDF task, so this run does not support replacing Doctrine.
Headline evidence
Strategy totals cover the identical subagent, intake, and PDF scenarios. Vision
was evaluated separately using the same generated image sentinel.
6.7×
Faster across the three StrategyOS scenarios
35.3s Grok vs 234.8s Doctrine
+48.5%
Higher aggregate strategy-run cost
$0.6089 Grok vs $0.4099 Doctrine
2 / 2
Successful end-to-end vision probes
Both returned “copper harbor” exactly
$1.2690
Total reported spend for all eight turns
Price caps remained in force
Lower is better in both charts. Bars use a shared scale within each chart;
labels show the exact observed value.
Latency by scenario
Seconds from turn request to terminal event
Vision · Grok4.9s
Vision · Doctrine11.4s
Subagents · Grok9.5s
Subagents · Doctrine43.8s
Strategy intake · Grok12.0s
Strategy intake · Doctrine36.6s
Sprinta PDF · Grok13.8s
Sprinta PDF · Doctrine154.3s
Grok 4.5Doctrine
Cost by scenario
Exact SDK/provider-reported USD cost
Vision · Grok$0.1297
Vision · Doctrine$0.1206
Subagents · Grok$0.2931
Subagents · Doctrine$0.1853
Strategy intake · Grok$0.1560
Strategy intake · Doctrine$0.0292
Sprinta PDF · Grok$0.1598
Sprinta PDF · Doctrine$0.1954
Grok 4.5Doctrine
Result ledger
| Scenario |
Model |
Outcome |
Latency |
Cost |
Observed behavior |
| Vision sentinel |
Grok 4.5 |
Pass |
4.9s |
$0.129679 |
Returned the exact visible text. |
| Vision sentinel |
Doctrine |
Pass |
11.4s |
$0.120580 |
Returned the exact visible text. |
| Two subagents |
Grok 4.5 |
Pass |
9.5s |
$0.293080 |
Started and completed both subagents; combined the words correctly. |
| Two subagents |
Doctrine |
Pass |
43.8s |
$0.185332 |
Started and completed both subagents; combined the words correctly. |
| Upmarket vs SMB intake |
Grok 4.5 |
Deferred |
12.0s |
$0.155999 |
Recognized the high-stakes decision and opened a four-question questionnaire. |
| Upmarket vs SMB intake |
Doctrine |
Deferred |
36.6s |
$0.029189 |
Framed the decision around PMF, unit economics, and constraints; asked three questions. |
| Sprinta PDF |
Grok 4.5 |
Mixed |
13.8s |
$0.159809 |
Read the attachment and asked one question, but the visible analysis was generic. |
| Sprinta PDF |
Doctrine |
Stronger |
154.3s |
$0.195368 |
Developed a concrete first-pass hypothesis and vulnerable assumptions before asking one question. |
Consulting-quality review
Speed and price are directly measured. Quality assessment is based on visible
assistant output and workflow behavior, not a separate model judge.
Grok 4.5
- Fastest model in every tested scenario.
- Correctly used StrategyOS, tools, subagents, questionnaires, PDF context, and image input.
- Produced concise terminal answers and avoided long overthinking.
- Did not demonstrate enough first-pass depth on the substantive PDF case.
“Broad company analysis needs StrategyOS first, then one scope lock so
the first-pass stays decision-grade rather than generic.”
Tegy Doctrine
- Slower in every scenario, dramatically so on the PDF case.
- Lower aggregate cost across the three StrategyOS scenarios.
- Produced the strongest substantive strategy work in this sample.
- Used more output on the PDF task before deferring to the user.
“Null hypothesis: Sprinta can validate its B2C high-touch coaching
subscription … and should therefore optimize this core funnel before
diverting resources to freemium plans or coach-infrastructure B2B.”
The questionnaire scenarios stopped at the deferred tool boundary. They prove
intake behavior, not final recommendation quality. A completed artifact comparison
requires identical answers to each model’s questions.
Vision finding
What was proven
Both models received the same generated PNG through Tegy’s real attachment
storage, sandbox materialization, Claude Agent SDK, Cloudflare AI Gateway,
and OpenRouter route. Both returned “copper harbor” exactly.
What remains unproven
This validates the end-to-end vision path and basic visual text recognition.
It is not a test of chart interpretation, slide critique, spatial reasoning,
or dense document-image analysis.
Published capability and price context
| Model |
OpenRouter ID |
Context |
Input modalities |
Input / output price |
| Grok 4.5 |
x-ai/grok-4.5 |
500K |
Text, image, file |
$2.00 / $6.00 per 1M |
| Doctrine |
moonshotai/kimi-k2.7-code |
262K |
Text, image |
$0.72 / $3.49 per 1M |
Sources:
SpaceXAI Grok 4.5 documentation,
OpenRouter xAI catalog, and
OpenRouter Kimi K2.7 Code.
Prices are a July 11, 2026 snapshot and may change.
Methodology and environment
This was a controlled local live-provider evaluation. The candidate mapping
existed only for the test and was removed immediately afterward.
- Runtime
- Claude Agent SDK 2.1.165 with StrategyOS and Tegy runtime skills loaded.
- Provider path
- Local Tegy container API → Cloudflare AI Gateway → OpenRouter.
- Effort
- Medium for both models.
- Cache status
- All captured model requests were cache misses.
- Cost safety
- $0.50 cap per vision turn and $2 cap per strategy turn.
- Cost source
- Terminal SDK/provider-reported
total_cost_usd.
Recommendation and next test
- Do not replace Doctrine from this evidence alone. Doctrine retained the strongest strategy-depth signal.
- Keep Grok 4.5 in evaluation as a fast deep-work candidate. The latency improvement is large enough to matter operationally.
- Run an end-to-end strategy artifact comparison. Supply identical business facts after intake and score the final recommendation, evidence use, decision clarity, and artifact quality.
- Add a real visual reasoning case. Use a chart, dashboard, or strategy slide—not only an OCR sentinel.
Verification and disclosure
npm run build passed before the live probes.
- Both live image attachment probes passed end to end.
- Both models completed all three selected StrategyOS scenarios without failure.
- The temporary Grok catalog mapping and vision-test selector were reverted.
- Internal chat IDs, turn IDs, session IDs, storage keys, credentials, and raw gateway captures are intentionally excluded.
- No video is included because this evaluates backend model behavior rather than a user-interface interaction.