Why this matters now
A year ago, the frontier model market had three meaningful options: GPT-4, Claude 3, and Gemini. Today, there are at least eight — spanning five price tiers from $1 to $30 per million output tokens.
The fragmentation is good for builders (more choices, lower prices) but paralyzing without a reference. Should you pay 30x more for Sol than Luna? When does Gemini 3.1 Pro beat Sonnet 5 on value? Is an open-weight model from China competitive with a US frontier lab on agentic coding?
This guide answers those questions with current pricing, published benchmarks, and a decision matrix. All data is sourced and current as of July 6, 2026.
The pricing landscape
| Model | Input / 1M tok | Output / 1M tok | Context | Category |
|---|---|---|---|---|
| Claude Opus 4.8 | $5.00 | $25.00 | 1M | Frontier |
| GPT-5.6 Sol | $5.00 | $30.00 | 128K | Frontier |
| Claude Sonnet 5 (std) | $3.00 | $15.00 | 200K | High |
| Claude Sonnet 5 (intro) | $2.00 | $10.00 | 200K | Promo |
| GPT-5.6 Terra | $2.50 | $15.00 | 128K | High |
| Gemini 3.1 Pro | $2.00 | $12.00 | 2M | High |
| Gemini 3.5 Flash | $1.50 | $9.00 | 1M | Mid |
| Mistral Medium 3.5 | $1.50 | $7.50 | 128K | Mid |
| Mistral Large 3 | $2.00 | $6.00 | 128K | Mid |
| GPT-5.6 Luna | $1.00 | $6.00 | 128K | Budget |
| GLM-5.2 | ~$0.73 | ~$2.29 | 1M | Budget |
| Kimi K2.7-Code | ~$0.95 | ~$4.00 | 256K | Budget |
The gap is 41x from cheapest input (GLM-5.2 at $0.73) to most expensive output (Sol at $30). But raw price per token is misleading — a model that costs 5x more but finishes the task in one attempt instead of three may be cheaper overall.
Key benchmarks
All scores are from official sources unless noted. SWE-bench Pro and Terminal-Bench 2.1 are the most relevant for agentic coding.
| Model | SWE-bench Pro | Terminal-Bench 2.1 | BrowseComp | OSWorld |
|---|---|---|---|---|
| Claude Opus 4.8 | 69.2% | 74.6% | — | 83.4% |
| Claude Sonnet 5 | 63.2% | 80.4% | 84.7% | 81.2% |
| GPT-5.6 Sol (Ultra) | — | 91.9% | — | — |
| GPT-5.6 Sol (base) | — | 88.8% | — | — |
| Gemini 3.1 Pro | ~58% | ~72% | — | — |
| Mistral Medium 3.5 | ~48% | ~62% | — | — |
| GLM-5.2 | 46.5% | 59.8% | 38.1% | — |
| Kimi K2.7-Code | 43.8% | 56.3% | 40.2% | — |
- Opus 4.8 leads on raw SWE-bench but Sonnet 5 beats it on Terminal-Bench 2.1 (agentic terminal use) and costs half as much
- GPT-5.6 Sol Ultra holds the highest single score (91.9% on Terminal-Bench) but is government-gated and costs $30/M output
- The open-weight models (GLM-5.2, Kimi K2.7) trail the frontier by 15–25 points on coding benchmarks but cost 5-10x less
- Gemini 3.1 Pro is competitive with Sonnet 5 on general reasoning but lags on agentic benchmarks
Cost-per-task analysis
Token pricing alone doesn’t tell you what a task actually costs. A coding agent using 50K input + 10K output tokens per iteration, averaging 3 iterations to complete a feature:
| Model | Cost per feature | Notes |
|---|---|---|
| Claude Opus 4.8 | $0.75 | Fast mode doubles output cost |
| Claude Sonnet 5 (std) | $0.45 | Best value in the high tier |
| Claude Sonnet 5 (intro) | $0.30 | Promo pricing won’t last |
| GPT-5.6 Sol | $0.90 | Most expensive per task |
| GPT-5.6 Terra | $0.45 | Same per-task as Sonnet 5 std |
| Gemini 3.1 Pro | $0.36 | 2M context a differentiator |
| Gemini 3.5 Flash | $0.27 | 1M context, fast |
| Mistral Medium 3.5 | $0.23 | Open-weight, self-hostable |
| GPT-5.6 Luna | $0.17 | Smart budget pick for simple tasks |
| GLM-5.2 | $0.08 | Cheapest by wide margin |
| Kimi K2.7-Code | $0.14 | Open-weight middle ground |
For a team running 500 agentic coding features per month:
- Opus 4.8: ~$375 — premium but highest autonomous success rate
- Sonnet 5: ~$225 — the sweet spot for most teams
- Gemini 3.1 Pro: ~$180 — competitive with Sonnet 5 if your task fits its strength profile
- Mistral Medium 3.5: ~$115 — open-weight, self-host, no vendor lock-in
- GLM-5.2: ~$40 — hard to beat on cost, but expect 15-25% higher retry rates
Decision matrix
| If you need… | Use… | Because |
|---|---|---|
| Highest autonomous coding success rate | Claude Opus 4.8 | Leads SWE-bench Pro at 69.2% |
| Best value for agentic coding | Claude Sonnet 5 | Near-Opus quality at half the price; beats Opus on Terminal-Bench |
| Lowest latency with good quality | GPT-5.6 Luna | Fastest tier, $1/$6, good for simple agent loops |
| Long-context reasoning (100K+ tokens) | Gemini 3.1 Pro | 2M context, best in class for document analysis |
| Open-weight, self-hosted deployment | Mistral Medium 3.5 or GLM-5.2 | Both run on your infra; Mistral has better benchmarks, GLM is cheaper |
| Government-gated high-effort reasoning | GPT-5.6 Sol Ultra | Highest single benchmark score (91.9%) but restricted |
| Maximum cost efficiency at scale | GLM-5.2 or Kimi K2.7-Code | 5-10x cheaper than frontier; good for high-volume, lower-stakes tasks |
| EU data sovereignty | Mistral Medium 3.5 (self-hosted) or Mistral Large 3 (API) | French company, Paris data center, sovereign by design |
The strategy: route, don’t commit
No single model wins every category. The right approach is a multi-provider gateway that routes tasks to the cheapest adequate model:
- Simple Q&A, classification, extraction → Luna / Mistral Medium 3.5 / GLM-5.2 ($0.08–$0.23/task)
- Coding with moderate complexity → Sonnet 5 / Terra / Gemini 3.1 Pro ($0.36–$0.45/task)
- Hard problems, autonomous multi-step → Opus 4.8 / Sol ($0.75–$0.90/task)
- Only when absolutely necessary → Sol Ultra (government-gated, $30/M output)
A team doing 80% of tasks on tier 1-2 and 20% on tier 3 averages ~$0.30/task — less than the cheapest model if they’d committed to a single provider.
What changed in the last month
- Sonnet 5 launch (Jun 30) collapsed the cost of near-Opus agentic quality — the single biggest pricing event this quarter
- GPT-5.6 family (Jun 26) introduced three-tier pricing, making OpenAI competitive again on cost with Luna/Terra
- Mistral Medium 3.5 went open-weight (May 22), giving the self-host market a credible mid-tier option
- GLM-5.2 (Jun 13) dropped input pricing to $0.73/M via third-party providers, 48% cheaper than at launch
- Gemini 3.1 Pro (May 19) locked Pro access behind paid API — free tier no longer includes frontier models
Related reading
- Claude Sonnet 5: Agentic Coding at Opus-Level for Half the Price
- GPT-5.6 Sol: Government-Gated at Launch
- Multi-Provider AI Gateways: Fallback Routing
- Mistral’s Full-Stack Pivot: Vibe, OCR 4, and Europe’s AI Platform Bet
- GLM-5.2: 1M Context and IndexShare Sparse Attention
- Kimi K2.7-Code: Open-Weights 1T MoE Coding Model
Sources
- Anthropic — Claude Sonnet 5 announcement
- Anthropic — Claude Opus 4.8 system card
- OpenAI — GPT-5.6 system card
- Google — Gemini 3.1 Pro technical report
- Mistral — Mistral Medium 3.5
- Z.ai — GLM-5.2 scorecard
- Moonshot AI — Kimi K2.7-Code model card
- Price Per Token — live model pricing tracker
About the author: Charles Jasthyn De La Cueva is an Admin Officer at PSU’s Quality Assurance Office and the builder behind ParSU-EDMS / QAOSYS. He writes about AI tools, infrastructure, and practical agent deployment.
