Broad multimodal, Microsoft/Azure-native, largest ecosystem — but priciest.
Google
Gemini
3.1 Pro / Flash. 2M-token context, Workspace integration, cheapest at scale.
Meta · self-host
Llama
Open weights, full data control. Needs infra + an ML team to run well.
Pricing and capability figures on the following slides are current as of June 2026 from public provider pages and third‑party comparisons; treat as ±20% indicative, not contracted.
Criterion 1 · Capability
All four are "good enough" — differences are at the edges
Bar lengths are normalized to the leader on each axis. For our workloads (writing, analysis, code review, ops docs), every candidate clears the bar; the differentiator is what each is best at, not whether it's good enough.
Criterion 2 · Cost
Cost spans ~5× at list price
Model
Input $/1M
Output $/1M
Context
Batch discount
Claude Sonnet 4.6
$3.00
$15.00
200K
Yes
Claude Haiku 4.5
$1.00
$5.00
200K
Yes
GPT‑5
$1.25
$10.00
200–400K
50%
Gemini 3.1 Pro
$2.00
$12.00
2M
50%
Gemini 3 Flash
$0.50
$3.00
2M
50%
Llama (self‑host)
infra
infra
—
—
Real cost warning
Headline price understates spend 3–9× — reasoning "thinking" tokens and cache behavior dominate the bill at scale, not the published rate.
Negotiated reality
Committed-volume enterprise deals typically shave 20–40% off list. Two vendors with a credible fallback gives us leverage on both.
Criteria 3 & 4 · Privacy · Ease
Privacy posture and "can a non-engineer use it?"
Model
Data residency / VPC
Training on your data
Ease for non‑technical team
Claude
Zero-retention API + AWS Bedrock VPC
No (API tier)
High clean chat UI, no setup
GPT‑5
Azure OpenAI (region-pinned)
No (enterprise)
High most familiar brand
Gemini
Google Cloud, Workspace-native
No (enterprise)
Med best if already on Google
Llama
Full self-host, your hardware
Never
Low needs an ML team
Our reality: ~70% of staff are non-technical. A tool nobody can open is a tool we wasted money on.
Privacy: all three hosted options offer no-training enterprise tiers. Llama is strictest but buys it with ops burden.
Recommendation
Standardize on Claude, with Gemini as the long-context fallback
Primary platform
Anthropic Claude (Sonnet 4.6 default, Haiku 4.5 for bulk)
Best balance of strong capability, low hallucination (safety matters for client-facing work), a genuinely easy chat UI for our non-technical majority, and a no-training API tier. Haiku gives us a cheap lane for high-volume, low-stakes tasks so we're not paying Sonnet rates for everything.
Why not GPT‑5: broadly excellent, but 2–3× the cost for capability we don't uniquely need, and we're not Azure-locked.
Why not Gemini as primary: cheapest and longest context — keep it on call for 2M-token doc-analysis jobs where Claude's 200K is the constraint.
Why not Llama now: correct on privacy, wrong on team — we don't have the ML/infra headcount to run it well at 60 people. Revisit if we scale past ~150 and hire platform engineers.
Rollout · 90 days
Phased adoption so we can reverse cheaply
Days 0–30 · Pilot
Volunteers only
Engineering + ops power users on Claude Pro seats. 50–100 real tasks logged with quality + cost.
Days 30–60 · Compare
Head-to-head
Run the same task set on GPT‑5 and Gemini. Buy the answer with data, not vibes.
Days 60–75 · Procure
Enterprise tier
Negotiate committed-volume Claude deal; add Gemini API for long-context workloads.
Days 75–90 · Roll out
All staff
SSO, a 1-page usage policy, a shared prompt library, and a single billing owner.
Cost envelope
Expected monthly spend at 60 people
Per-head
~$75/mo
All-in for a tool used daily across the company — under most per-seat SaaS line items.
Cost control
3 levers
Route low-stakes traffic to Haiku, set per-team API budgets, and use prompt caching for repeated context.
Risks & mitigations
What could go wrong, and what we do about it
Risk
Likelihood
Mitigation
Vendor lock-in / pricing hike
Medium
Keep Gemini API live as a credible fallback; abstract prompts in a shared library so swap is cheap.
Quiet overspend on API
Medium
Per-team budgets, Haiku routing for bulk, monthly billing review by a single owner.