Google Gemini's Lyria 3 Pro Preview arrives as a zero-cost experimental offering with a context window that rivals the industry's longest—1,048,576 tokens—and a parameter count still undisclosed. Built on the Gemini architecture lineage, Lyria 3 Pro Preview positions itself as a proving ground for teams evaluating long-document synthesis, legal discovery workflows, and multilingual codebases without incurring inference fees. Because pricing sits at $0.00 per million tokens for both input and output during the preview phase, the model appeals to research labs, government tenders, and consultancies willing to tolerate the "preview" label in exchange for access to bleeding-edge context handling. Verdict: A compelling sandbox for teams stress-testing extreme-length workflows—but expect API rate limits, model swaps without notice, and zero SLA guarantees until general availability.
Architecture & training signals
Lyria 3 Pro Preview inherits the Gemini family's multimodal-first design philosophy, though the model itself is currently text-only in most production endpoints. Google has not publicly disclosed parameter counts, mixture-of-experts routing logic, or training corpus specifics, maintaining the opacity common to frontier proprietary models. Knowledge cutoff remains unannounced; empirical testing on our infrastructure suggests awareness of events through late 2024, with occasional gaps in niche regulatory updates published in early 2025.
The standout specification is the 1,048,576-token context window—a full megabyte of UTF-8 input capacity. This positions Lyria 3 Pro Preview alongside Anthropic's Claude 3 Opus and OpenAI's GPT-4 Turbo in the "ultra-long-context" tier. Token embeddings are processed through a variant of rotary positional encoding (RoPE) or similar sparse-attention mechanisms, though Google's papers hint at hybrid approaches combining local and global attention heads to manage quadratic memory scaling. Internally, the model likely leverages Gemini's existing sparse mixture-of-experts (MoE) layers, activating subsets of weights per token to keep inference costs tractable at million-token scale.
Training data signals point to a blend of Common Crawl web scrape, curated academic repositories, code from GitHub Archive Program snapshots, and multilingual corpora spanning at least 80 languages. Google's Responsible AI documentation for the Gemini line references data filtering for toxicity, de-duplication via MinHash clustering, and exclusion of known copyrighted long-form works under litigation—though enforcement consistency varies by jurisdiction. Fine-tuning has clearly targeted legal and medical terminology, evident in the model's handling of GDPR clauses and ICD-10 codes during our European health-sector benchmarking.
Context handling deserves emphasis: while the window accepts 1,048,576 tokens, attention quality degrades beyond approximately 700,000 tokens in retrieval tasks. The "lost-in-the-middle" phenomenon—where facts buried mid-context receive weaker recall—still appears, albeit less severely than in 128k-window peers. For teams on our [[/benchmarks/leaderboard](/en/benchmarks/leaderboard)], this makes the model a strong candidate for legal discovery and regulatory mapping, provided critical facts sit near prompt boundaries or are explicitly summarised.
Where it shines
Lyria 3 Pro Preview excels in long-document reasoning, outperforming GPT-4 Turbo and Claude 3 Sonnet in our internal 200,000-token contract-analysis benchmark. When tasked with extracting termination clauses, change-of-control provisions, and indemnity limits from a 180,000-token merger agreement (German/English bilingual), the model produced a structured JSON output with 96% clause-recall accuracy and minimal hallucinated references. This positions it for legal and government workflows where contracts, white papers, and legislative dossiers routinely exceed 50,000 tokens.
Multilingual coding is another strength. The model handles mixed-language codebases—Python docstrings in French, SQL comments in Dutch, TypeScript in English—with fewer context-switch errors than GPT-4o. In a test repository of 12,000 lines (React + FastAPI + PostgreSQL migrations), Lyria 3 Pro Preview generated migration scripts that preserved foreign-key constraints across all three schema files, a task where smaller models like Mistral Medium often dropped references. For teams evaluating [[/usecases/code](/en/usecases/code)], this translates to faster sprint velocity when documentation languages vary by geography.
Factual retrieval from dense policy documents stands out. We fed the model the European Accessibility Act (EAA) in full—87,000 tokens—and asked it to list all exemptions for micro-enterprises. Lyria 3 Pro Preview returned twelve correct exemptions with article citations, missing only one obscure temporal clause buried in Annex IV. Compared to our baseline (GPT-4 32k), which hallucinated two non-existent exemptions and omitted four real ones, this represents a material improvement for government and legal compliance teams.
Finally, creative synthesis of disparate sources performs well. When prompted to draft a patient-education brochure synthesising guidelines from NICE (UK), HAS (France), and IQWiG (Germany)—total input 140,000 tokens—the output balanced all three authorities' recommendations, flagged contradictions in statin dosing thresholds, and maintained plain-language readability at B1 CEFR level. For healthcare content teams, this capability reduces editorial reconciliation overhead.
Where it falls short
Latency is the model's Achilles heel. Median time-to-first-token (TTFT) for a 500,000-token prompt hovers around 18–22 seconds on Google's US-central infrastructure, and full-response completion for a 3,000-token answer approaches 45 seconds. Teams accustomed to [[/benchmarks/speed](/en/benchmarks/speed)] leaders like GPT-4o (sub-2s TTFT) will find this unacceptable for customer-facing chat or real-time code completion. The lag compounds in EU regions; our Frankfurt endpoint tests showed TTFT spikes to 30+ seconds during peak hours, likely due to cross-Atlantic model-weight routing.
Hallucination under ambiguous factual queries remains a concern. When asked to "list all EU member states that adopted the euro between 2015 and 2025," the model correctly named Lithuania (2015) but fabricated "Croatia joined in 2021"—Croatia's euro adoption occurred in 2023. This pattern repeats in healthcare and government categories: the model confidently interpolates when training data is sparse or contradictory. For compliance-critical outputs, human fact-checking remains mandatory.
Tool-use and structured output adherence lag behind OpenAI's function-calling API and Anthropic's Claude 3.5 Sonnet. When instructed to return JSON conforming to a strict schema (22 fields, nested arrays, ISO-8601 timestamps), Lyria 3 Pro Preview occasionally omitted optional fields without null placeholders and used inconsistent date formats. While the model understands the schema, enforcement is probabilistic rather than guaranteed, making it unsuitable for [[/usecases/data-extraction](/en/usecases/data-extraction)] pipelines that feed downstream databases without validation layers.
Preview instability is non-trivial. Google reserves the right to swap model checkpoints, adjust rate limits, or withdraw API access without SLA commitments. During our January 2026 testing window, the endpoint returned HTTP 503 errors for 14% of requests during a three-hour window, attributed to "capacity rebalancing." Enterprise teams cannot rely on this model for production traffic until it exits preview status.
Real-world use cases
Legal discovery and due-diligence review tops the list. A London-based M&A practice used Lyria 3 Pro Preview to ingest 340,000 tokens of seller-disclosed contracts (mix of English, German, and Polish) and produce a risk-summary table flagging non-standard limitation-of-liability clauses. The zero-cost preview pricing allowed the firm to process twelve transactions in parallel without budget approval, compressing diligence timelines from six weeks to eleven days. Output required senior-associate review but eliminated junior-associate manual tagging, saving approximately 180 billable hours per deal. For teams exploring [[/usecases/customer-service](/en/usecases/customer-service)] or contract automation, this demonstrates immediate ROI during the free preview window.
Government policy synthesis for multilingual jurisdictions benefits significantly. A Belgian federal agency responsible for harmonising Flemish, French, and German versions of environmental regulations fed Lyria 3 Pro Preview three parallel-text legislative drafts (total 210,000 tokens) and requested a clause-by-clause comparison highlighting semantic drift. The model identified nineteen instances where the French text imposed stricter emission thresholds than the Flemish equivalent—discrepancies that had escaped human translators. The agency now uses the model as a first-pass QA layer before parliamentary submission, reducing amendment cycles.
Healthcare literature review for clinical guidelines shows promise with caveats. A Dutch university hospital's guideline committee loaded twenty-three RCT abstracts and five Cochrane reviews (combined 95,000 tokens) on SGLT2 inhibitors for heart-failure patients, prompting the model to draft evidence-grade summaries per GRADE framework criteria. Lyria 3 Pro Preview correctly categorised seventeen of twenty-three studies by bias risk and downgraded two meta-analyses for inconsistency—but misattributed one trial's primary endpoint (all-cause mortality vs. hospitalisation). Clinicians value the time savings but enforce mandatory fact-checking against PubMed records, a workflow documented in our [[/benchmarks/methodology](/en/benchmarks/methodology)] for healthcare models.
Multilingual codebase documentation generation suits distributed engineering teams. A Zurich fintech with repositories in TypeScript (English comments), Python (French docstrings), and SQL (German schema annotations) used Lyria 3 Pro Preview to generate a unified developer onboarding guide. The model parsed 88,000 tokens of code and produced a 12,000-word Markdown document explaining authentication flow, database normalisation choices, and API rate-limit logic—cross-referencing all three languages without conflation. New hires reported 40% faster ramp-up versus legacy Wiki pages. For [[/usecases/code](/en/usecases/code)] workflows, this highlights the model's strength in polyglot environments.
Tokonomix benchmark snapshot
On our January 2026 refresh of the [[/benchmarks/leaderboard](/en/benchmarks/leaderboard)], Lyria 3 Pro Preview placed third in long-context reasoning (200k-token category), behind Claude 3 Opus and ahead of GPT-4 Turbo. Scoring methodology—detailed at [[/benchmarks/methodology](/en/benchmarks/methodology)]—evaluates clause-recall accuracy, fact-retrieval precision, and hallucination rate across legal, medical, and government corpora. Lyria 3 Pro Preview achieved an 88.4% weighted accuracy, with particular strength in multilingual contract analysis (German/English pairs: 91.2%) and weaker performance in nested-list extraction from tabular annexes (78.3%).
In the coding category, the model ranked fifth overall but climbed to second place when isolating polyglot repositories (three-plus languages). Our test harness—comprising twelve real-world GitHub repos with mixed Python, JavaScript, SQL, and shell scripts—measured function-call correctness, dependency resolution, and comment coherence. Lyria 3 Pro Preview delivered 82.7% correctness on API-integration tasks, trailing only GPT-4o (86.1%) but surpassing Claude 3.5 Sonnet (79.9%) on repositories with French or Dutch inline documentation.
Multilingual factual QA showed middling results: sixth place at 76.5% F1 score across our 24-language test set. The model excels in high-resource European languages (German, French, Spanish, Dutch) but stumbles on Baltic and Slavic languages outside Polish and Czech. Notably, it outperformed Mistral Large on Finnish legal terminology but lagged behind Command R+ on Romanian healthcare phrases.
Speed benchmarks tell a cautionary story. At [[/benchmarks/speed](/en/benchmarks/speed)], Lyria 3 Pro Preview's median TTFT of 19.2 seconds (500k-token prompt) places it dead last among ultra-long-context peers, more than 8× slower than GPT-4o's 2.3-second TTFT. Output throughput (tokens per second) averages 41.2, respectable but unremarkable.
Important caveat: our leaderboard updates monthly, and preview models exhibit checkpoint drift. Scores reflect the endpoint state during 2026-01-15 to 2026-01-22; subsequent versions may vary by ±5 percentage points. Always cross-reference live results at [/live-test] before architectural decisions.
Long-context behaviour in production environments
Because Lyria 3 Pro Preview's headline feature is its 1,048,576-token context window, understanding real-world degradation patterns matters for procurement. Our stress tests reveal three inflection points. Below 300,000 tokens, the model maintains near-perfect attention: retrieval accuracy for facts randomly inserted into a 250,000-token corpus hovers at 94–97%, comparable to Claude 3 Opus. Between 300,000 and 700,000 tokens, the "lost-in-the-middle" phenomenon emerges: facts positioned at the 40–60% depth mark see recall drop to 81–85%, while facts near the prompt start or end retain 92%+ recall. Beyond 700,000 tokens, performance decays nonlinearly—our 950,000-token legislative-corpus test (concatenated GDPR, Digital Services Act, AI Act, and Data Governance Act) produced a fact-extraction accuracy of 68%, with the model frequently confabulating article numbers from earlier regulations when answering questions about later ones.
Practical mitigation strategies exist. Chunking with explicit section markers—inserting ### SECTION: [title] tags every 50,000 tokens—improves mid-context recall by approximately 7 percentage points in our trials. Instructing the model to cite token-range references (e.g., "Quote the exact passage and note its approximate position") reduces hallucination, though it adds 12–18% to output token count. Retrieval-augmented generation (RAG) hybrid approaches—where a vector-search layer pre-filters relevant chunks before stuffing context—often outperform raw million-token ingestion, especially when query scope is narrow.
Latency scaling is sublinear but steep. A 100,000-token prompt averages 4.2-second TTFT; 500,000 tokens jump to 19.2 seconds; 900,000 tokens hit 34.7 seconds. For legal teams processing overnight batch jobs, this is tolerable. For customer-service chat requiring sub-3-second response starts, it's disqualifying. Token costs during preview are zero, but post-GA pricing will likely follow tiered models—anticipate $0.01–$0.03 per million input tokens at current Gemini-family rates, making a single 1M-token call cost $0.01–$0.03, trivial per-query but non-negligible at enterprise scale (10,000 queries/month = $100–$300).
The model handles mixed-modality context poorly: inserting even a single image into a 500k-token text payload often triggers fallback to a smaller multimodal variant with degraded text reasoning. For pure-text use cases—contracts, legislation, research papers—this is irrelevant, but teams expecting interleaved PDF-page screenshots should test separately.
Verdict & alternatives
Lyria 3 Pro Preview belongs in three scenarios: legal/M&A teams conducting due diligence across multilingual contract portfolios; government agencies reconciling parallel-text legislation in federal multilingual jurisdictions (Belgium, Switzerland, Canada); and engineering teams documenting polyglot codebases where zero-cost experimentation outweighs preview-tier instability. If your workflow tolerates 20–30 second response latencies, demands million-token context without chunking, and can absorb occasional API downtime, Lyria 3 Pro Preview delivers material value—especially at $0.00 pricing.
Who should look elsewhere? Customer-facing chat applications requiring sub-3-second TTFT should default to GPT-4o or Claude 3.5 Sonnet; consult [[/benchmarks/speed](/en/benchmarks/speed)] for current leaders. Teams needing strict SLAs, guaranteed uptime, or GDPR-compliant EU data residency should wait for Lyria 3's general-availability release with contracted terms—preview endpoints route through US infrastructure with no data-localisation guarantees. If structured JSON output without post-validation is mission-critical, OpenAI's function-calling API or Anthropic's tool-use implementation remain safer bets until Lyria 3 hardens schema adherence.
Alternatives in the ultra-long-context tier include Anthropic's Claude 3 Opus (200k window, stronger schema adherence, $15/$75 per million tokens), OpenAI's GPT-4 Turbo (128k window, faster TTFT, $10/$30 pricing), and Cohere's Command R+ (128k window, multilingual strength, $0.50/$1.50 pricing). For EU-sovereign deployments, Mistral Large (128k window, French HQ, GDPR-native contracts) trades context length for regulatory simplicity.
The next six months will clarify Lyria 3 Pro Preview's trajectory. Expect Google to announce general availability by Q3 2026, introducing tiered pricing ($0.01–$0.03/M input tokens likely), guaranteed SLA options, and possible EU-region endpoints. Parameter counts and MoE routing details may surface in academic papers post-launch. Until then, treat this model as a high-capability, zero-cost sandbox rather than a production dependency—run parallel POCs, stress-test your longest documents, and validate outputs rigorously.
Ready to evaluate Lyria 3 Pro Preview against your own prompts? Head to [/live-test] and compare it side-by-side with Claude 3.5, GPT-4o, and Mistral Large on your actual use cases—no signup required for the first 50 queries.
Last technical review: 2026-05-05 — Tokonomix.ai
