SLP AI Report Card
We gave free AI models the same 10 school speech-language pathology tasks, from a vocalic /r/ word list to an IEP progress-report template, and scored every answer 1–5 on accuracy, clinical soundness, therapy-readiness, SLP specificity and format. Every output is published in full, so you can check our grading.
Tested 2026-10-164 models · 10 tasks · 40 scored outputsHow we score →
Biggest finding: no model wrote a correct vocalic /r/ word list. Every model either included words without the "er" sound (erode, garden, puddle), misplaced word positions (calling "bird" word-final), looped, or returned nothing. The best score on that task was 2.8 out of 5. Check every AI word list against the actual target sound before using it in therapy.
Watch the facts in parent and legal documents: all four sound-development handouts had wrong age norms or cited studies we couldn't verify. On evaluation templates, Qwen3.8 27B put 85–115 at the "5th–95th percentile" (it's the 16th–84th), and GPT-OSS 120B named the wrong IDEA category ("Specific Learning Disability – Speech/Language" instead of Speech or Language Impairment). On the other hand, GPT-OSS 120B and Qwen3.8 27B both wrote strong, neurodiversity-affirming pragmatics plans and clean progress-report templates.
Note on Qwen3.8 27B: its first 5 tasks ran with a 3,000-token budget. Groq's free tier then started rejecting those requests (Qwen's free output limit is 1,000 tokens per minute), so its other 5 tasks ran with a 1,500-token budget, and 2 of them were cut off. Each output page lists the budget it ran with.
Rankings
| # | Model | Overall | Best area | Free tier | Privacy | |
|---|---|---|---|---|---|---|
| 1 | GPT-OSS 120B OpenAI (open-weight) · open model via Groq | 4.4/5 | Social pragmatics (4.8) | Groq free tier: 30 requests/min, 1,000 requests/day, 8K tokens/min, 200K tokens/day. source | Groq says it does not retain inference inputs/outputs by default; it may log them for up to 30 days for reliability or abuse checks. Never paste student names or personal info. source | Scores & full outputs → |
| 2 | Qwen3.8 27B Qwen (Alibaba) · open model via Groq | 4.0/5 | Social pragmatics (5.0) | Groq free tier: 30 requests/min, 1,000 requests/day, 8K tokens/min. source | Groq says it does not retain inference inputs/outputs by default; it may log them for up to 30 days for reliability or abuse checks. Never paste student names or personal info. source | Scores & full outputs → |
| 3 | GPT-OSS 20B OpenAI (open-weight) · open model via Groq | 3.6/5 | Social pragmatics (4.8) | Groq free tier: 30 requests/min, 1,000 requests/day, 8K tokens/min, 200K tokens/day. source | Groq says it does not retain inference inputs/outputs by default; it may log them for up to 30 days for reliability or abuse checks. Never paste student names or personal info. source | Scores & full outputs → |
| 4 | ALLaM 2 7B SDAIA · open model via Groq | 2.2/5 | IEP & documentation (2.8) | Ran on our Groq free-tier account; Groq's published limits table does not list this model. source | Groq says it does not retain inference inputs/outputs by default; it may log them for up to 30 days for reliability or abuse checks. Never paste student names or personal info. source | Scores & full outputs → |
Open models were tested through the free Groq API (one run per task, temperature 0.3). Every score is one run per task: a snapshot, not a verdict.
Not yet tested
| Tool | Maker | Status |
|---|---|---|
| ChatGPT (free) | OpenAI | Not yet tested. We only rank tools we've actually run the full task set on. We never estimate a score. |
| Gemini (free) | Not yet tested. We only rank tools we've actually run the full task set on. We never estimate a score. | |
| Claude | Anthropic | Not yet tested. We only rank tools we've actually run the full task set on. We never estimate a score. |
| Microsoft Copilot | Microsoft | Not yet tested. We only rank tools we've actually run the full task set on. We never estimate a score. |
Want to see a tool tested? Tell us.