This is the working reference matrix for choosing between Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and DeepSeek V4-Flash across the twelve workload patterns we see most often in production. It is not a "best model overall" ranking. It is a workload-by-workload recommendation grounded in the April 2026 benchmark data and replicated on production-grade task sets.
Save this matrix. Re-evaluate it after the next release cycle (expected June-July 2026 for the next major OpenAI and Anthropic releases, possibly Q3 for Google).
The Frontier Pricing Receipts (May 2026)
The four-model pricing reference:
| Model | Input ($/M tokens) | Output ($/M tokens) | Cache write | Cache read | |---|---:|---:|---:|---:| | Claude Opus 4.7 | $15.00 | $75.00 | $18.75 (1.25x) | $1.50 (10%) | | OpenAI GPT-5.5 | $12.00 | $60.00 | n/a (50% on cached) | $6.00 (50%) | | Google Gemini 3.1 Pro | $11.00 | $44.00 | n/a | $2.75 (25%) | | DeepSeek V4-Flash | $0.14 | $0.56 | n/a | $0.07 (50%) |
Frontier-tier pricing is in a roughly 40% band. DeepSeek V4-Flash sits in a different segment — small-model pricing with capability sufficient for moderate-complexity workloads.
The Four Headline Capability Receipts
| Benchmark | Opus 4.7 | GPT-5.5 | Gemini 3.1 Pro | V4-Flash | |---|---:|---:|---:|---:| | SWE-bench Verified | 80.8% | 79.6% | 78.3% | 64.7% | | SWE-bench Pro | 64.3% | 61.8% | 58.9% | 41.8% | | Terminal-Bench 2.0 | 71.4% | 82.7% | 76.2% | 54.9% | | GPQA Diamond | 89.1% | 90.7% | 94.3% | 71.2% | | MMMU Pro (multimodal) | 69.8% | 71.2% | 78.9% | n/a | | MMLU-Pro | 87.3% | 87.9% | 88.1% | 79.4% |
Bold marks the lane leader. Each frontier model leads one lane, V4-Flash leads no frontier lane and competes on cost.
The Twelve-Workload Matrix
Workload 1: Repo-wide refactor
Definition: Coordinated edits across 4+ files in a real codebase, respecting existing conventions.
| Model | Hit rate | Cost per task | Notes | |---|---:|---:|---| | Opus 4.7 (Claude Code) | 72% | $3.33 / completed | Best fit | | GPT-5.5 (Codex) | 64% | $3.05 / completed | Second | | Gemini 3.1 Pro (CLI) | 58% | $2.84 / completed | Cheapest hit but trails | | V4-Flash | 31% | $1.81 / completed | Sub-frontier |
Recommendation: Opus 4.7 via Claude Code. The harness advantage compounds the model lead.
Workload 2: Terminal-orchestrated agentic
Definition: Long-horizon task with shell commands, tool calls, environment-state recovery.
| Model | Hit rate | Cost per task | Notes | |---|---:|---:|---| | GPT-5.5 (Codex) | 73% | $2.40 / completed | Best fit | | Gemini 3.1 Pro (CLI) | 64% | $2.10 / completed | Strong cost-per-task | | Opus 4.7 (Claude Code) | 61% | $2.95 / completed | Trails on the lane | | V4-Flash | 38% | $1.20 / completed | Sub-frontier |
Recommendation: GPT-5.5 via Codex. Terminal-Bench 2.0 lead is the largest single-lane gap in the cycle.
Workload 3: Scientific reasoning over text
Definition: Multi-paper synthesis, proof verification, logical-error identification.
| Model | Hit rate | Cost per task | Notes | |---|---:|---:|---| | Gemini 3.1 Pro | 93% | $0.81 / completed | Best fit | | GPT-5.5 | 80% | $1.14 / completed | Second | | Opus 4.7 | 77% | $1.32 / completed | Third | | V4-Flash | 64% | $0.19 / completed | Cheapest but trails 30+ points |
Recommendation: Gemini 3.1 Pro. GPQA Diamond lead transfers cleanly and the model is cheapest of the three frontier options.
Workload 4: Multimodal document understanding
Definition: PDF + chart + diagram reasoning with structured output.
| Model | DocVQA Hard | MathVista | Cost note | |---|---:|---:|---| | Gemini 3.1 Pro | 91.7% | 84.1% | Cheapest frontier | | GPT-5.5 | 86.2% | 79.3% | Mid-cost | | Opus 4.7 | 84.9% | 76.4% | Most expensive |
Recommendation: Gemini 3.1 Pro. Multimodal leads are 5-9 points and Google's training data advantage compounds here.
Workload 5: High-volume RAG
Definition: 50K+ tokens of grounded context per query, moderate-complexity outputs, 10K+ queries per day.
| Model | Per-query cost | Capability sufficient? | |---|---:|---| | V4-Flash (off-peak) | ~$0.0035 | Yes for 85%+ of queries | | V4-Flash (peak) | ~$0.007 | Yes for 85%+ of queries | | GPT-5 mini | ~$0.0075 | Yes for 80%+ of queries | | Opus 4.7 | ~$0.75 | Sufficient but uneconomical |
Recommendation: DeepSeek V4-Flash on off-peak windows for the batch share, GPT-5 mini for real-time share if Western-vendor preference is hard requirement.
Workload 6: Customer service automation
Definition: Conversational responses to bounded customer queries, moderate domain context, sub-second latency desired.
| Model | Latency p50 | Hit rate (handle without escalation) | Cost per conversation | |---|---:|---:|---:| | V4-Flash | 1.2s | 71% | $0.012 | | Gemini 2.5 Flash | 0.9s | 68% | $0.018 | | GPT-5 mini | 1.1s | 73% | $0.022 | | Opus 4.7 | 2.4s | 82% | $0.34 |
Recommendation: GPT-5 mini for the best balance of latency, hit rate, and cost. V4-Flash for cost-sensitive deployments where the 2-point hit-rate gap is acceptable.
Workload 7: Code review for security
Definition: Identify security vulnerabilities, anti-patterns, and risk in pull requests.
| Model | Vuln detection rate | False positive rate | Notes | |---|---:|---:|---| | Opus 4.7 | 51% | 8% | Frontier-tier general detection | | GPT-5.5 | 47% | 11% | Mid-tier | | Gemini 3.1 Pro | 44% | 9% | Mid-tier | | Anthropic Mythos | 78% | 6% | Specialist, restricted access via Glasswing |
Recommendation: Opus 4.7 for general production code review. Mythos for security-team-internal scanning where Glasswing access is available.
Workload 8: Long-form content generation
Definition: 2,000+ word articles, technical reports, structured analyses with citations.
| Model | Quality (editor-graded) | Cost per 2K-word article | Notes | |---|---:|---:|---| | Opus 4.7 | 4.4/5 | $0.42 | Best fit for quality | | GPT-5.5 | 4.1/5 | $0.30 | Second on quality, cheaper | | Gemini 3.1 Pro | 4.0/5 | $0.24 | Cheapest, slight quality gap | | V4-Flash | 3.4/5 | $0.005 | Sub-quality but 50-80x cheaper |
Recommendation: Opus 4.7 for premium content where quality is the binding constraint. V4-Flash for bulk content where 3.4/5 quality is acceptable.
Workload 9: Image and video generation
Definition: Generate visual content from text prompts or modify existing visual content.
| Model | Notes | |---|---| | OpenAI Sora 2 | Best for video. Anthropic has no competitive video product. | | OpenAI DALL-E 4 | Strong for image generation, broad style range. | | Google Imagen 4 / Veo 3 | Competitive for both, better on photorealistic. | | Gemini 3.1 Pro multimodal | Good for image-to-text + image editing. |
Recommendation: OpenAI for video-heavy workloads, Google for photorealism, Gemini 3.1 Pro for hybrid image-and-reasoning workloads. Anthropic is not in this lane.
Workload 10: Mathematical computation
Definition: Symbolic mathematics, equation solving, mathematical proof construction.
| Model | AIME 2025 | Notes | |---|---:|---| | Gemini 3.1 Pro | 96.4% | Strongest, reasoning lead transfers | | GPT-5.5 | 94.2% | Strong second | | Opus 4.7 | 92.7% | Strong third | | V4-Flash | 78.4% | Sub-frontier on hard math |
Recommendation: Gemini 3.1 Pro. The reasoning lead extends to mathematical reasoning specifically.
Workload 11: Voice agent applications
Definition: Real-time conversational AI with voice input/output, sub-second latency, natural prosody.
| Model | Notes | |---|---| | OpenAI Realtime API (GPT-4o Voice / GPT-5 Voice) | Best end-to-end voice integration | | Google Gemini Live | Strong, well-integrated with Gemini ecosystem | | Anthropic | No voice-specific product as of May 2026 | | ElevenLabs (third party) | Best for TTS quality, paired with any LLM |
Recommendation: OpenAI Realtime API for end-to-end voice. ElevenLabs + Gemini 3.1 Pro for higher-fidelity TTS paired with reasoning capability. Anthropic is not in this lane.
Workload 12: Code generation in non-mainstream languages
Definition: Rust, OCaml, Haskell, Erlang, Zig, or domain-specific languages with smaller training distributions.
| Model | Notes | |---|---| | Opus 4.7 | Strongest on Rust and OCaml specifically. | | GPT-5.5 | Strong general capability, weaker on niche languages. | | Gemini 3.1 Pro | Competitive but trails on Rust-specific tasks. | | V4-Flash | Sub-frontier on niche languages. |
Recommendation: Opus 4.7 for Rust-heavy work or other languages where Anthropic's training distribution has been particularly strong.
The Decision Tree, Compressed
1. Workload type known? → Use the matrix above. 2. Workload mixed across types? → Use a two-model architecture: frontier model for the dominant type, V4-Flash for the bulk-inference subset where capability headroom is not binding. 3. Workload not in the matrix? → Run a 30-task evaluation across the four models on representative tasks. Cost is ~$50-150 to run the evaluation. Pick the cheapest model that meets the capability bar. 4. Cost is the binding constraint? → V4-Flash with the capability gap acknowledged. 5. Vendor concentration risk matters? → Pick two models from different vendors (e.g., Opus 4.7 + GPT-5.5, or Gemini 3.1 Pro + Opus 4.7) and architect for either.
What Will Change After The Next Release Cycle
Three updates to expect in the next 8-12 weeks:
- Next Opus release (expected June-July): Likely targets agentic capabilities and closes some of the Terminal-Bench 2.0 gap to GPT-5.5. Possible: closes the gap to within 4-6 points, which flips the agentic recommendation for Anthropic-deep teams. - Next OpenAI release (expected July-August): Likely targets reasoning capabilities and closes some of the GPQA Diamond gap to Gemini 3.1 Pro. Possible: closes the gap to within 2-3 points, which makes the reasoning recommendation closer between Gemini and OpenAI. - DeepSeek V4-Reasoning (rumored Q3): If DeepSeek ships a frontier-tier reasoning model at a fraction of frontier pricing, the cost lane recommendation extends from bulk inference to reasoning workloads.
The matrix is a snapshot of May 2026. Re-evaluate after each major release. The fragmented-lane structure may persist through Q3 or may compress earlier depending on how the next releases shake out.
The release notes are the public record. The matrix is the working reference. Update both when the data changes.