The ORCFLO Indexis an independent benchmark that evaluates large language models the way business professionals actually use them — across real-world tasks spanning analysis, writing, extraction, summarization, and behavioral reliability. Each model is scored on three dimensions (quality, cost, and speed) by a panel of four independent judges. This report evaluates Codestral (2508) in the context of 32 models from ORCFLO Index — May 10, 2026 cohort · 32 total models tested · Anthropic, Google, OpenAI, Mistral.
Codestral (2508)
Codestral (2508) finishes #32 of 32 on overall quality (71.09), placing it last in the cohort behind frontier leaders like GPT 5 (95.04) and Gemini 3 Pro (93.17), and trailing even close peers GPT 4.1 Nano (71.3) and GPT 4o Mini (72.8). Its redeeming traits are operational: #1 in speed at 2.4 seconds and #8 in cost at $0.0006 per case. Treat it as a latency-optimized utility model, not a reasoning engine.
Key Findings
- Last-place overall quality at 71.09 (#32 of 32), with bottom-of-field ranks in Analysis (#32), Writing (#32), Hallucination (#32), and Refusal Calibration (#31).
- Fastest model in the cohort at 2.4 seconds average response time (#1 of 32), roughly 5x faster than o3-mini (12.1s) and notably ahead of peers like GPT 4o Mini (4.8s).
- Summarization is the lone bright spot at 85.5 (#24), the only category where Codestral reaches contender tier rather than trailing tier.
- Cost efficiency is mid-pack, not best-in-class at $0.0006 per case (#8); Mistral Small 3 ($0.0001) and GPT 4.1 Nano ($0.0002) deliver higher quality at lower price.
Model Performance: Quality & Cost
The chart below plots quality against cost for all 32 models in the ORCFLO Index. Each dot represents the average quality score a model achieved across the full basket of real-world business tasks, alongside the cost in credits to complete the entire test suite. Models in the upper-left quadrant deliver the highest quality at the lowest cost.
Model Performance: Quality & Time Elapsed
Quality alone doesn’t tell the full story — response time determines whether a model is viable for time-sensitive workflows. The chart below plots each model’s quality score against the total time required to complete the test suite. Models in the upper-left deliver the best quality with the least delay.
Category Scorecard
The ORCFLO Indexevaluates models using real-world business tasks — not academic puzzles or synthetic benchmarks. Each test case is designed to expose specific differences in how models handle the work professionals actually do. Scores are averaged across each category and ranked independently across all 32 models.
| Category | Score | Rank | Tier |
|---|---|---|---|
| Abilities — Core language tasks: what the model can produce when given a well-formed prompt. | |||
| AnalysisReasoning, strategic judgment, disqualifying-factor detection | #32 | Trailing | |
| ExtractionField accuracy, null handling, format compliance, zero fabrication | #29 | Trailing | |
| SummarizationCompression quality, key-point retention, length compliance | #24 | Contender | |
| WritingTone, structure, persuasion, audience adaptation | #32 | Trailing | |
| Behaviors — How the model acts under pressure: reliability, compliance, and restraint. | |||
| HallucinationFabrication detection, factual grounding, source fidelity | #32 | Trailing | |
| Instruction FollowingConstraint adherence, format compliance, multi-part directives | #30 | Trailing | |
| Refusal CalibrationAppropriate refusal vs. over-refusal on legitimate requests | #31 | Trailing | |
| Stability — Repeatability and predictability across identical inputs. | |||
| Output ConsistencyRun-to-run reproducibility, format stability, score variance | #31 | Trailing | |
Strengths and Cautions
Strengths
- Fastest response time in the field at 2.4 seconds (#1 of 32) — useful for latency-sensitive pipelines where sub-3-second turnaround matters more than depth.
- Summarization holds at 85.5 (#24, contender tier), the model's strongest single category and the only one not in trailing tier.
- Low absolute cost at $0.0006 per case (#8 of 32), keeping it viable for high-volume, low-stakes workloads.
Cautions
- Analysis and Writing both rank dead last at 52.5 and 68.2 respectively (#32 in each) — disqualifying for any task involving reasoning, synthesis, or polished prose.
- Hallucination rank of #32 (72.0) signals the highest factual risk in the cohort; not suitable for customer-facing or compliance-sensitive outputs.
- Refusal Calibration at 56.9 (#31) and Output Consistency at 74.1 (#31) indicate unreliable behavior on edge cases and reproducibility-dependent workflows.
Head-to-Head: Frontier Models
Codestral (2508) is Mistral’s speed-first budget option in the ORCFLO Index. The table below compares it against the top-performing models from each major provider. Tier assignments use 25% quartiles across the full 32-model field.
| Model | Quality Avg | Quality Rank | Cost Rank | Speed Rank |
|---|---|---|---|---|
| Mistral Small 3 (24B) | 79.9 | #26 | #1 | #7 |
| Gemini 2.0 Flash-Lite | 78.0 | #27 | #3 | #2 |
| o3-mini | 77.4 | #28 | #27 | #20 |
| GPT 4o | 75.4 | #29 | #16 | #6 |
| GPT 4o Mini | 72.8 | #30 | #6 | #8 |
| GPT 4.1 Nano | 71.3 | #31 | #2 | #3 |
| Codestral (2508) | 71.1 | #32 | #8 | #1 |
When to Use Codestral (2508)
The ORCFLO Index
This evaluation covers 40 cases across 8 categories. All tasks are text-only and English-only. Code generation, multi-turn conversation, multimodal tasks, and agentic workflows are not tested. Each contestant is scored by a panel of four independent judges — Gemini 2.5 Pro, Claude Opus 4.7, GPT 5.5, and Mistral Large — with final scores averaged across all four. Cost and speed measurements reflect API pricing and latency as of the test date (May 10, 2026) and will change as providers update their offerings.
How We Test
The ORCFLO Indexevaluates large language models across three independent dimensions — quality, cost, and speed — using real-world business tasks designed to expose the differences that matter for model selection. Each model is scored by a panel of four independent judges to reduce single-model bias.
- Test Cases
- 40 cases across 8 categories spanning Abilities (Analysis, Extraction, Summarization, Writing), Behaviors (Hallucination, Instruction Following, Refusal Calibration), and Stability (Output Consistency).
- Judge Panel
- Gemini 2.5 Pro, Claude Opus 4.7, GPT 5.5, and Mistral Large. Each judge scores independently. Final score is the average across all four.
- Scoring
- Three independent ranks: quality (higher is better), cost (lower is better), speed (faster is better). No composite score — composites hide the tradeoffs that drive model-selection decisions.
- Tier Definitions
- LeaderQuality ≥ 90.8Ranks 1–8Strong≥ 85.9Ranks 9–16Contender≥ 80.9Ranks 17–24Trailing< 80.9Ranks 25–32