What Fable 5.1 Cracking a 370-Year Cipher Means for Research Workflows

Vals AI reports that Claude Fable 5.1 solved the Cyphral Distich, a 370-year-old cipher. The headline is fun if you like historical puzzles. The practical implication is bigger: frontier reasoning models are getting good enough to work through messy evidence, generate testable hypotheses, use long context, and produce an audit trail that humans can inspect.

That matters because most high-value knowledge work is not a clean question-answer task. Archival research, legal discovery, fraud investigation, intelligence analysis, incident response, compliance review, patent analysis, and technical due diligence all involve incomplete documents, contradictory clues, uncertain metadata, and repeated verification. A cipher is an extreme benchmark for the same pattern: gather fragments, infer structure, test explanations, reject weak paths, and document the strongest interpretation.

This post is a practical guide for builders and research teams. You’ll learn what changed, what workflows are now possible, how to design verification loops, where to use premium reasoning models, and how to route cheaper models for extraction and preprocessing. Cost is the proof layer: the right architecture can reserve expensive frontier reasoning for the small part of the job where it matters.

---

What changed: from chatbot answers to auditable reasoning loops

The important shift is not that a model “knows” the answer to an old cipher. It is that a frontier model can operate in a workflow where the answer must be earned. The model has to inspect source material, propose candidate interpretations, reason across fragments, preserve uncertainty, and use tools or human review to verify each claim.

That is the same architecture modern research teams need. A document investigation workflow has four stages:

1. Ingestion: OCR, parse, chunk, classify, and extract entities from messy files.
2. Hypothesis generation: Identify patterns, gaps, inconsistencies, or possible explanations.
3. Tool-assisted testing: Run searches, compare sources, calculate timelines, check citations, and test alternate interpretations.
4. Verification and reporting: Produce evidence-backed conclusions with confidence levels and citations.

The Cyphral Distich news is timely because the frontier model layer is improving exactly where old automation failed: ambiguity. Traditional OCR, search, and rules-based extraction can find text, dates, names, and matches. They struggle when documents are partial, language is archaic, names are misspelled, metadata is wrong, or clues only make sense after several reasoning steps.

Long context also changes the operating model. With 1,000,000-token context windows available in models like Claude Fable 5.1, Claude Mythos 5.1, GPT-5.2, and GPT-5.6 Sol, teams can place entire research packets into a single run: source excerpts, extracted entities, timelines, competing hypotheses, prior notes, and verification checklists.

[stat] 1,000,000 tokens Context available in Claude Fable 5.1, GPT-5.2, GPT-5.6 Sol, Claude Opus 5, and several other frontier models — enough for large evidence packets and multi-document review.

The market cares because investigation work is expensive and slow. Legal teams pay for document review. Enterprises pay analysts to trace incidents. Banks investigate fraud rings. Researchers spend weeks reconciling archives. If a model can accelerate the middle 60% of that work while producing auditable reasoning, the ROI is immediate.

---

7 workflows builders can create now

The best way to think about this capability is not “AI solves ciphers.” It is “AI runs structured uncertainty workflows over messy evidence.” Here are seven practical systems teams can build.

1. Archival research assistant

Build a system that ingests scanned letters, handwritten notes, OCR output, catalog descriptions, and prior scholarship. The model extracts entities, dates, places, relationships, and uncertain readings, then builds a research memo with citations.

Useful for:

• Historical societies digitizing collections
• University research teams
• Museums and archives
• Journalists working with historical records
• Genealogy and provenance research

Premium model role: final synthesis, uncertain interpretation, cross-source reasoning.

Cheaper model role: OCR cleanup, entity extraction, language normalization, metadata tagging.

2. Cipher and codebook analysis workspace

A cipher-solving workflow is a strong test case for hypothesis loops. The system can generate candidate substitution patterns, test frequency distributions, compare known language corpora, and maintain a ranked list of possible decryptions.

Useful for:

• Historical cryptography
• Puzzle and challenge platforms
• Intelligence-style training
• Red-team reasoning benchmarks
• Educational tools

Premium model role: infer structure, generate unconventional hypotheses, evaluate competing explanations.

Cheaper model role: run candidate transformations, summarize failed attempts, extract repeated sequences.

3. Legal archive and discovery review

Legal archives often include email threads, contracts, scanned exhibits, deposition transcripts, regulatory filings, and handwritten notes. A reasoning workflow can build timelines, identify contradictions, flag missing evidence, and prepare attorney review packets.

Useful for:

• Litigation support
• Regulatory investigations
• Contract disputes
• Internal investigations
• FOIA document review

Premium model role: contradiction analysis, theory-of-case synthesis, risk memo drafting.

Cheaper model role: document classification, clause extraction, timeline extraction.

4. Fraud evidence graph builder

Fraud review depends on connecting weak signals: repeated addresses, shared devices, timing anomalies, inconsistent identities, transaction patterns, and communications. A model-assisted evidence graph can extract relationships and propose investigative leads.

Useful for:

• Fintech fraud teams
• Insurance investigations
• Marketplace trust and safety
• Procurement fraud
• AML support workflows

Premium model role: causal hypothesis generation and narrative reconstruction.

Cheaper model role: entity extraction, duplicate detection, transaction summarization.

5. Technical incident investigation

Security incidents, outages, and production failures are evidence problems. Logs, tickets, deploy histories, Slack threads, metrics, and traces have to be reconciled into a timeline and root cause hypothesis.

Useful for:

• Site reliability engineering
• Security operations
• Post-incident reviews
• Vendor incident analysis
• Compliance evidence preparation

Premium model role: root-cause reasoning, alternative hypothesis testing, executive summary.

Cheaper model role: log summarization, ticket clustering, event extraction.

6. Patent and prior-art investigation

Prior-art research requires comparing claims, dates, diagrams, technical language, and product documentation. Long-context models can read many documents and produce claim charts with citations.

Useful for:

• IP litigation
• Patent prosecution
• Technical diligence
• R&D landscaping
• Competitive intelligence

Premium model role: claim interpretation and similarity reasoning.

Cheaper model role: document retrieval, technical term extraction, abstract summarization.

7. Investigative journalism evidence desk

Reporters work with leaks, court filings, public databases, interviews, financial documents, and inconsistent naming. A tool-assisted reasoning system can maintain evidence trails and separate confirmed facts from leads.

Useful for:

• Newsroom investigations
• Nonprofit accountability research
• Public records analysis
• Sanctions and ownership tracing
• Local government oversight

Premium model role: story hypothesis, contradiction analysis, source reliability review.

Cheaper model role: OCR cleanup, source summaries, named entity extraction.

---

Workflow 1: Archival research packet with verification loop

This workflow is for teams working with historical documents, court archives, government records, or research collections. The goal is not to replace a researcher. The goal is to turn a messy corpus into a structured, auditable research packet.

Step 1: Ingest and normalize documents

Start with raw PDFs, scans, OCR text, catalog metadata, and notes. Run OCR and convert every source into a structured object:

Field	Example
document_id	archive_box_12_letter_004
source_type	letter, ledger, court filing, catalog note
date_raw	“Michaelmas term, 1656”
date_normalized	1656-09 to 1656-12
author	uncertain: “J. Wilkins?”
text	OCR or transcription
confidence	OCR confidence, catalog confidence
citations	page, scan URL, archive location

Use a cheaper model such as Gemini 2.5 Flash-Lite, GPT-5 nano, Mistral Small 3.2, or DeepSeek V4.1 Flash for normalization. This stage is repetitive and should not use your most expensive model.

Step 2: Extract entities and uncertain readings

Ask the extraction model to produce structured JSON:

• People
• Organizations
• Places
• Dates
• Uncertain words
• Cross-references
• Document condition notes
• Possible aliases or spelling variants

Prompt pattern:

Extract entities and uncertainties from the document.
Return JSON only.
Preserve original spellings.
For every extracted claim, include the exact source quote and character span when available.
Mark uncertain readings as low, medium, or high confidence.

The key is to preserve uncertainty instead of forcing false precision. Historical work often turns on ambiguous handwriting, archaic spelling, damaged pages, or calendar differences.

Step 3: Build a timeline and evidence graph

Use extracted fields to build a timeline. Link entities across documents, but keep confidence scores. “J. Wilkins,” “John Wilkins,” and “Dr. Wilkins” may be the same person, but the graph should show why.

At this stage, a mid-tier long-context model can help. GPT-5.2 costs $1.75 input / $14 output per 1M tokens and supports 1,000,000 context. Gemini 3 Flash costs $0.50 / $3 per 1M tokens with 1,000,000 context. Use these for timeline assembly before escalating to premium reasoning.

Step 4: Generate hypotheses

Now route to a frontier reasoning model such as Claude Fable 5.1, Claude Mythos 5.1, GPT-5.6 Sol, or GPT-5.2 pro. Ask it to produce competing hypotheses, not a single answer.

Prompt pattern:

You are reviewing an archival evidence packet.
Generate 3-5 competing hypotheses that explain the documents.
For each hypothesis:
- list supporting evidence with citations
- list contradicting evidence with citations
- identify missing evidence that would change the conclusion
- assign a confidence score from 0-100
- propose the next 5 verification steps
Do not treat uncertain OCR or inferred metadata as confirmed fact.

Step 5: Verify claims with tools

Run searches against your archive index, external catalog APIs, citation databases, or internal document store. Every claim should be checked against source text. The model should produce verification tasks that tools execute.

Examples:

• Search for spelling variants of a name
• Compare dates against known historical calendars
• Check whether a place name existed at the claimed time
• Retrieve all documents mentioning the same seal, signature, or address
• Validate citations and page references

Step 6: Produce an auditable memo

The final memo should separate:

• Confirmed facts
• Strong inferences
• Weak hypotheses
• Rejected hypotheses
• Open questions
• Source citations
• Verification log

This is where premium models earn their cost. You are paying for structured synthesis, uncertainty management, and reasoning over conflicting evidence.

---

Workflow 2: Fraud or legal evidence investigation loop

This workflow adapts the same reasoning pattern to commercial investigations: fraud, litigation, compliance, employment investigations, vendor disputes, procurement abuse, or incident review.

Step 1: Define the investigation question

Good workflows start with a specific question:

• “Did these accounts coordinate to exploit a promotion?”
• “Which vendor invoices are linked to the same operator?”
• “What evidence supports breach of contract?”
• “Which internal controls failed before the incident?”
• “Is there a pattern of backdated approvals?”

Write the investigation question into the system prompt and require the model to keep its answer scoped.

Step 2: Create an evidence schema

Before feeding documents to a model, define what counts as evidence.

Evidence type	Fields to extract
Transaction	amount, timestamp, account, counterparty, payment method
Communication	sender, recipient, timestamp, quoted commitment, attachment
Identity	name, address, device, email, IP, phone
Contract	party, clause, obligation, deadline, exception
Event	actor, action, time, system, source log
Risk signal	anomaly, rule triggered, model score, analyst note

Extraction can run on inexpensive models. DeepSeek V4.1 Flash costs $0.15 input / $0.60 output per 1M tokens. Gemini 2.0 Flash-Lite costs $0.075 / $0.30 per 1M tokens. GPT-5 nano costs $0.05 / $0.40 per 1M tokens.

Step 3: Build a contradiction matrix

Ask a mid-tier model to identify conflicts:

• Same person, different address
• Same device, different account owner
• Invoice date before purchase approval
• Contract obligation missed before termination notice
• Account activity during impossible travel window
• Internal email contradicting public statement

Prompt pattern:

Review the extracted evidence table.
Find contradictions, impossible sequences, repeated identifiers, and suspicious timing.
Return a contradiction matrix with:
- contradiction_id
- evidence A citation
- evidence B citation
- why the two conflict
- severity from 1-5
- recommended verification step

Step 4: Run frontier reasoning on the top evidence bundle

Do not send every raw document to a premium model. Send the top contradictions, relevant source excerpts, timelines, entity graph, and investigation question.

Use Claude Fable 5.1 when the reasoning burden is high and the evidence is messy. It costs $10 input / $50 output per 1M tokens with 1,000,000 context. Use Claude Opus 5 at $5 / $25 per 1M tokens when you want premium Anthropic reasoning at half the Fable 5.1 token rate. Use GPT-5.6 Sol at $5 / $30 per 1M tokens for strong frontier synthesis with the same 1,050,000 context family as other GPT-5.6 models.

The frontier prompt should require adversarial review:

You are an investigation reasoning model.
Given the evidence bundle, produce:
1. The strongest explanation supported by evidence
2. The strongest innocent explanation
3. Evidence that would falsify each explanation
4. Claims that are not supported
5. A ranked list of next verification actions
6. A final confidence score and why it is not higher
Every factual claim must cite evidence_id values.

Step 5: Human review and final report

The final output should never be a black-box accusation. It should be a review packet:

• Executive summary
• Timeline
• Evidence table
• Contradiction matrix
• Model-generated hypotheses
• Human reviewer notes
• Verification status
• Recommended action

This format works for legal review because it preserves provenance. It works for fraud teams because it shows why a case was escalated. It works for compliance because it produces an audit trail.

---

Model Choice and Cost

The main cost principle is simple: do not use the most expensive reasoning model for extraction. Premium models should be reserved for the short, high-value steps: hypothesis generation, contradiction reasoning, final synthesis, and adversarial verification.

Here is a practical routing table.

Workflow stage	Recommended model	Price per 1M input/output tokens	Why use it
OCR cleanup and simple extraction	GPT-5 nano	$0.05 / $0.40	Very cheap for repetitive parsing
Bulk summarization	Gemini 2.0 Flash-Lite	$0.075 / $0.30	Lowest-cost high-volume summarization
Entity extraction and tagging	DeepSeek V4.1 Flash	$0.15 / $0.60	Cheap extraction with 1M context
Timeline assembly	Gemini 3 Flash	$0.50 / $3	Long context with moderate cost
Mid-tier synthesis	GPT-5.2	$1.75 / $14	Strong general reasoning, 1M context
Premium investigation reasoning	Claude Opus 5	$5 / $25	Premium reasoning at lower price than Fable 5.1
Frontier cipher-style reasoning	Claude Fable 5.1	$10 / $50	Best fit for hard ambiguous reasoning loops
Premium OpenAI alternative	GPT-5.6 Sol	$5 / $30	Strong long-context synthesis
Maximum pro reasoning	GPT-5.2 pro	$21 / $168	Use only for highest-stakes final review

Example cost: archival packet review

Assume one research packet uses:

• 600,000 input tokens for extracted documents, notes, timeline, and source excerpts
• 25,000 output tokens for hypotheses, verification plan, and memo

Cost by model:

Model	Input cost	Output cost	Total per run
DeepSeek V4.1 Flash	$0.09	$0.02	$0.11
Gemini 3 Flash	$0.30	$0.08	$0.38
GPT-5.2	$1.05	$0.35	$1.40
Claude Opus 5	$3.00	$0.63	$3.63
Claude Fable 5.1	$6.00	$1.25	$7.25
GPT-5.2 pro	$12.60	$4.20	$16.80

The right answer is not always the cheapest run. The right answer is a routed workflow: use DeepSeek, Gemini Flash, GPT-5 nano, or Mistral Small for extraction; use Fable 5.1, Opus 5, GPT-5.6 Sol, or GPT-5.2 pro for the final reasoning pass.

Example cost: 1,000 investigation packets

Using the same 600,000 input / 25,000 output packet size:

Model	Cost per run	Cost per 1,000 runs
DeepSeek V4.1 Flash	$0.11	$105
Gemini 3 Flash	$0.38	$375
GPT-5.2	$1.40	$1,400
Claude Opus 5	$3.63	$3,625
Claude Fable 5.1	$7.25	$7,250
GPT-5.2 pro	$16.80	$16,800

Use AI Cost Check to adjust these numbers for your own token counts. If you are comparing OpenAI and Anthropic options, start with GPT-5 vs Claude Opus 4.6 or GPT-5 vs Gemini 3 Pro to sanity-check model families and token economics.

---

When premium reasoning is worth it

Premium models are worth the cost when the task has ambiguity, high stakes, and a real cost of being wrong. The Cyphral Distich is a symbolic example because ciphers require long chains of reasoning and rejection of plausible false paths. The same pattern appears in business and research.

Use a premium model when:

• The evidence is incomplete or contradictory
• Multiple explanations fit the facts
• The final report will influence legal, financial, or reputational decisions
• You need structured uncertainty, not just a summary
• The model must reason across hundreds of pages
• Human experts will review and challenge the result
• A wrong answer costs more than the model bill

Do not use a premium model when:

• The task is pure extraction
• The source documents are clean and repetitive
• You only need a short summary
• A rules engine can solve it
• The workflow lacks citations and verification
• You cannot tolerate probabilistic outputs
• There is no human review for high-stakes conclusions

For many teams, Claude Opus 5 is the more economical premium default at $5 / $25 per 1M tokens. Claude Fable 5.1 at $10 / $50 should be reserved for the hardest ambiguous reasoning cases: cipher-like investigations, fragile evidence synthesis, adversarial review, or final memos where the reasoning quality is worth the extra cost.

If you want a cheaper OpenAI stack, use GPT-5 nano or GPT-5 mini for preprocessing, GPT-5.2 for mid-tier synthesis, and GPT-5.6 Sol for final review. If you need a low-cost open-family route, DeepSeek V4.1 Flash, Mistral Small 3.2, and Llama 4 Scout are strong candidates for bulk document stages.

---

Architecture pattern: the investigation router

A production system should route tasks by difficulty. The router is more important than the model choice because it prevents runaway costs and improves auditability.

Layer 1: Ingestion services

Use OCR, parsers, and file converters. Store raw text, page images, source metadata, and confidence scores. Never overwrite originals.

Layer 2: Cheap extraction models

Run low-cost models for JSON extraction, language cleanup, summarization, and entity tagging. Store every extracted field with citations.

Recommended models:

• GPT-5 nano: $0.05 / $0.40
• Gemini 2.0 Flash-Lite: $0.075 / $0.30
• DeepSeek V4.1 Flash: $0.15 / $0.60
• Mistral Small 3.2: $0.10 / $0.30

Layer 3: Retrieval and evidence graph

Build a vector index, keyword index, entity graph, and timeline database. Retrieval should return exact excerpts, not paraphrases. Use embeddings and deterministic filters for dates, names, document IDs, and source types.

Layer 4: Mid-tier synthesis

Use models like GPT-5.2, Gemini 3 Flash, or Mistral Large 3 to generate timelines, contradiction matrices, and initial hypotheses.

Layer 5: Frontier reasoning

Only send compressed evidence bundles to premium models. The bundle should include:

• Investigation question
• Timeline
• Entity graph summary
• Top source excerpts
• Contradictions
• Prior hypotheses
• Verification checklist
• Required report format

Layer 6: Verification tools

The model should call tools or produce machine-readable tasks:

• Retrieve source excerpt
• Search archive
• Compare dates
• Validate citation
• Check calculation
• Run alternate cipher transform
• Query case database
• Recompute graph relationships

Layer 7: Human signoff

High-stakes workflows require a human reviewer. The final UI should show source excerpts beside each claim. Confidence scores should be explained, not hidden.

---

Risks, limits, and guardrails

The biggest risk is false confidence. Frontier models can produce elegant explanations from weak evidence. That is dangerous in legal, fraud, compliance, and historical interpretation. A good workflow must make unsupported claims visible.

Hallucinated citations

Require document IDs and exact quotes for every factual claim. If the model cannot cite a claim, the claim should be labeled “uncited” and excluded from the final conclusion.

Over-compression

Summaries can erase important caveats. Preserve raw excerpts and allow reviewers to inspect the original source. For legal and historical work, never rely only on a summary layer.

Confirmation bias

If you ask the model to prove a theory, it will find supporting patterns. Require the model to generate the strongest opposing explanation and list falsifying evidence.

Privacy and privilege

Legal, HR, medical, and fraud investigations may include sensitive data. Use appropriate data controls, retention policies, access logs, and model providers approved for your compliance environment.

Cost runaway

Long-context models make it easy to send too much data. Track tokens per stage. Add routing thresholds. Cap retries. Use cheap models for document-level work and premium models for final reasoning.

Benchmark mismatch

Solving a historical cipher does not prove a model can handle your regulated workflow. Build evaluation sets from your own documents. Include known answers, edge cases, noisy OCR, and adversarial examples.

---

Practical prompts for builders

Use these as starting templates.

Evidence extraction prompt

You are extracting structured evidence from a source document.
Return JSON only.

Rules:
- Preserve original wording.
- Include source quotes for every extracted field.
- Mark uncertainty explicitly.
- Do not infer facts not stated in the document.

Schema:
{
  "document_id": "...",
  "entities": [],
  "dates": [],
  "events": [],
  "claims": [],
  "uncertain_readings": [],
  "source_quotes": []
}

Hypothesis generation prompt

You are analyzing an evidence packet.
Generate competing hypotheses, not a single answer.

For each hypothesis:
- summary
- supporting evidence IDs
- contradicting evidence IDs
- missing evidence
- confidence score
- verification steps
- what would falsify it

Separate confirmed facts from interpretation.

Final audit memo prompt

Produce an audit-ready memo.

Sections:
1. Question investigated
2. Confirmed facts
3. Timeline
4. Competing hypotheses
5. Recommended conclusion
6. Evidence table
7. Unsupported claims excluded
8. Verification log
9. Human review checklist

Every factual sentence must include evidence IDs.

These prompts work best when paired with structured data and retrieval tools. A raw pile of PDFs will produce weaker results than a curated evidence packet.

---

Recommended stacks by team type

Small research team

Use:

• Gemini 2.0 Flash-Lite for OCR cleanup and summaries
• DeepSeek V4.1 Flash for extraction
• GPT-5.2 for synthesis
• Claude Opus 5 for final review

This keeps costs low while preserving a premium reasoning step.

Legal or compliance team

Use:

• GPT-5 nano for bulk extraction
• Gemini 3 Flash for long-context timeline assembly
• Claude Fable 5.1 for hard contradiction analysis
• Human attorney or compliance reviewer for signoff

Prioritize audit logs, citations, and access control.

Fraud operations team

Use:

• DeepSeek V4.1 Flash for entity and transaction extraction
• Graph database for relationships
• GPT-5.2 for contradiction matrices
• Claude Opus 5 or GPT-5.6 Sol for escalated case narratives

Route only top-risk cases to premium models.

Frontier research lab

Use:

• Multiple cheap models for independent extraction
• Deterministic scripts for cipher transforms or statistical checks
• Claude Fable 5.1 for hypothesis search
• GPT-5.2 pro for adversarial second opinion
• A benchmark suite with known-answer cases

This is the closest analog to the Cyphral Distich use case.

---

Frequently asked questions

What does Fable 5.1 solving the Cyphral Distich mean for builders?

It means frontier reasoning models are becoming useful for structured investigation workflows where evidence is messy, incomplete, and ambiguous. Builders should copy the pattern: cheap extraction, long-context evidence packets, premium hypothesis generation, tool-assisted verification, and human signoff.

How much does an archival research workflow cost?

A large archival packet with 600,000 input tokens and 25,000 output tokens costs about $0.11 on DeepSeek V4.1 Flash, $1.40 on GPT-5.2, $3.63 on Claude Opus 5, or $7.25 on Claude Fable 5.1. Use AI Cost Check to calculate your own packet size.

Should I use Claude Fable 5.1 for every document?

No. Use cheaper models like GPT-5 nano, Gemini 2.0 Flash-Lite, Mistral Small 3.2, or DeepSeek V4.1 Flash for extraction and summarization. Reserve Claude Fable 5.1 for final reasoning, ambiguous hypotheses, and verification-heavy synthesis.

What is the best model stack for legal or fraud evidence review?

Use a routed stack: cheap extraction with GPT-5 nano or DeepSeek V4.1 Flash, long-context timeline assembly with Gemini 3 Flash or GPT-5.2, and final reasoning with Claude Opus 5, Claude Fable 5.1, or GPT-5.6 Sol. This architecture keeps the audit trail strong while avoiding premium-model spend on repetitive parsing.

How do I prevent hallucinations in investigation workflows?

Require exact citations, source quotes, confidence labels, and a verification log for every factual claim. The final report should separate confirmed facts from hypotheses and include the strongest opposing explanation before recommending a conclusion.

---

Build the workflow, then price the route

The Cyphral Distich result is a useful market signal: frontier reasoning is moving from impressive demos into practical investigation systems. The opportunity for builders is to package that capability into workflows that handle uncertainty, preserve evidence, and produce reviewable conclusions.

Start with one narrow use case: an archive packet, a fraud case, a legal timeline, a post-incident review, or a prior-art comparison. Build the pipeline with cheap extraction first. Add retrieval and verification. Then reserve Claude Fable 5.1, Claude Opus 5, GPT-5.6 Sol, or GPT-5.2 pro for the reasoning steps where the extra cost changes the quality of the decision.

Use AI Cost Check to model your per-run and monthly spend, compare frontier options, and test fallback routes before you ship. If you are choosing between model families, review GPT-5 vs Gemini 3 Pro, GPT-5 vs DeepSeek V3.2, and Claude Opus 4.6 vs DeepSeek V3.2 for broader cost-performance context.